Skip to content

M
MiniCPM-training
Commit beaacea5 by wyt2000
Options

feat: support pretrain for minicpm

parent 2d2620bd
Showing with 3594 additions and 1248 deletions
pretrain/LLaMA-Factory/.gitignore
......@@ -167,4 +167,5 @@ config/
saves/
output/
wandb/
data/
submit.sh
ret_one
pretrain/LLaMA-Factory/configs/minicpm_pretrain.yaml 0 → 100644
### model
model_name_or_path: /lustre/S/huangdi/open_for_out/models/MiniCPM-training/models/MiniCPM-quant
### method
stage: pt
do_train: true
finetuning_type: full
### dataset
dataset: c4_demo
cutoff_len: 2048
overwrite_cache: true
preprocessing_num_workers: 80
# streaming: true
# max_steps: 102500
# should_process: false
### output
output_dir: saves/bitnet-cpt-on-the-stack-all-32x40g-r8a100b-0726
logging_steps: 1
save_steps: 3000
plot_loss: true
overwrite_output_dir: true
### train
# resume_from_checkpoint: saves/bitnet-cpt-on-the-stack-all-32x40g-r8a100b-0726/checkpoint-75000
per_device_train_batch_size: 8
gradient_accumulation_steps: 1
learning_rate: 5.0e-5
num_train_epochs: 1.0
lr_scheduler_type: cosine
warmup_steps: 375
adam_beta1: 0.9
adam_beta2: 0.95
bf16: true
ddp_timeout: 180000000
cache_dir: /lustre/S/huangdi/open_for_out/models/MiniCPM-training/pretrain/LLaMA-Factory/cache
pretrain/LLaMA-Factory/data/README_zh.md 0 → 100644
[dataset_info.json](dataset_info.json) 包含了所有可用的数据集。如果您希望使用自定义数据集，请**务必**在 `dataset_info.json` 文件中添加*数据集描述*，并通过修改 `dataset: 数据集名称` 配置来使用数据集。
目前我们支持 **alpaca** 格式和 **sharegpt** 格式的数据集。
```json
"数据集名称": {
"hf_hub_url": "Hugging Face 的数据集仓库地址（若指定，则忽略 script_url 和 file_name）",
"ms_hub_url": "ModelScope 的数据集仓库地址（若指定，则忽略 script_url 和 file_name）",
"script_url": "包含数据加载脚本的本地文件夹名称（若指定，则忽略 file_name）",
"file_name": "该目录下数据集文件夹或文件的名称（若上述参数未指定，则此项必需）",
"formatting": "数据集格式（可选，默认：alpaca，可以为 alpaca 或 sharegpt）",
"ranking": "是否为偏好数据集（可选，默认：False）",
"subset": "数据集子集的名称（可选，默认：None）",
"split": "所使用的数据集切分（可选，默认：train）",
"folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
"num_samples": "该数据集所使用的样本数量。（可选，默认：None）",
"columns（可选）": {
"prompt": "数据集代表提示词的表头名称（默认：instruction）",
"query": "数据集代表请求的表头名称（默认：input）",
"response": "数据集代表回答的表头名称（默认：output）",
"history": "数据集代表历史对话的表头名称（默认：None）",
"messages": "数据集代表消息列表的表头名称（默认：conversations）",
"system": "数据集代表系统提示的表头名称（默认：None）",
"tools": "数据集代表工具描述的表头名称（默认：None）",
"images": "数据集代表图像输入的表头名称（默认：None）",
"chosen": "数据集代表更优回答的表头名称（默认：None）",
"rejected": "数据集代表更差回答的表头名称（默认：None）",
"kto_tag": "数据集代表 KTO 标签的表头名称（默认：None）"
},
"tags（可选，用于 sharegpt 格式）": {
"role_tag": "消息中代表发送者身份的键名（默认：from）",
"content_tag": "消息中代表文本内容的键名（默认：value）",
"user_tag": "消息中代表用户的 role_tag（默认：human）",
"assistant_tag": "消息中代表助手的 role_tag（默认：gpt）",
"observation_tag": "消息中代表工具返回结果的 role_tag（默认：observation）",
"function_tag": "消息中代表工具调用的 role_tag（默认：function_call）",
"system_tag": "消息中代表系统提示的 role_tag（默认：system，会覆盖 system column）"
}
}
```
## Alpaca 格式
### 指令监督微调数据集
- [样例数据集](alpaca_zh_demo.json)
在指令监督微调时，`instruction` 列对应的内容会与 `input` 列对应的内容拼接后作为人类指令，即人类指令为 `instruction\ninput`。而 `output` 列对应的内容为模型回答。
如果指定，`system` 列对应的内容将被作为系统提示词。
`history` 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮对话的指令和回答。注意在指令监督微调时，历史消息中的回答内容**也会被用于模型学习**。
```json
[
{
"instruction": "人类指令（必填）",
"input": "人类输入（选填）",
"output": "模型回答（必填）",
"system": "系统提示词（选填）",
"history": [
["第一轮指令（选填）", "第一轮回答（选填）"],
["第二轮指令（选填）", "第二轮回答（选填）"]
]
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"columns": {
"prompt": "instruction",
"query": "input",
"response": "output",
"system": "system",
"history": "history"
}
}
```
### 预训练数据集
- [样例数据集](c4_demo.json)
在预训练时，只有 `text` 列中的内容会用于模型学习。
```json
[
{"text": "document"},
{"text": "document"}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"columns": {
"prompt": "text"
}
}
```
### 偏好数据集
偏好数据集用于奖励模型训练、DPO 训练和 ORPO 训练。
它需要在 `chosen` 列中提供更优的回答，并在 `rejected` 列中提供更差的回答。
```json
[
{
"instruction": "人类指令（必填）",
"input": "人类输入（选填）",
"chosen": "优质回答（必填）",
"rejected": "劣质回答（必填）"
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"ranking": true,
"columns": {
"prompt": "instruction",
"query": "input",
"chosen": "chosen",
"rejected": "rejected"
}
}
```
### KTO 数据集
- [样例数据集](kto_en_demo.json)
KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人类反馈。
```json
[
{
"instruction": "人类指令（必填）",
"input": "人类输入（选填）",
"output": "模型回答（必填）",
"kto_tag": "人类反馈 [true/false]（必填）"
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"columns": {
"prompt": "instruction",
"query": "input",
"response": "output",
"kto_tag": "kto_tag"
}
}
```
### 多模态数据集
- [样例数据集](mllm_demo.json)
多模态数据集需要额外添加一个 `images` 列，包含输入图像的路径。目前我们仅支持单张图像输入。
```json
[
{
"instruction": "人类指令（必填）",
"input": "人类输入（选填）",
"output": "模型回答（必填）",
"images": [
"图像路径（必填）"
]
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"columns": {
"prompt": "instruction",
"query": "input",
"response": "output",
"images": "images"
}
}
```
## Sharegpt 格式
### 指令监督微调数据集
- [样例数据集](glaive_toolcall_zh_demo.json)
相比 alpaca 格式的数据集，sharegpt 格式支持**更多的角色种类**，例如 human、gpt、observation、function 等等。它们构成一个对象列表呈现在 `conversations` 列中。
注意其中 human 和 observation 必须出现在奇数位置，gpt 和 function 必须出现在偶数位置。
```json
[
{
"conversations": [
{
"from": "human",
"value": "人类指令"
},
{
"from": "function_call",
"value": "工具参数"
},
{
"from": "observation",
"value": "工具结果"
},
{
"from": "gpt",
"value": "模型回答"
}
],
"system": "系统提示词（选填）",
"tools": "工具描述（选填）"
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"formatting": "sharegpt",
"columns": {
"messages": "conversations",
"system": "system",
"tools": "tools"
}
}
```
### 偏好数据集
- [样例数据集](dpo_zh_demo.json)
Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的消息，并在 `rejected` 列中提供更差的消息。
```json
[
{
"conversations": [
{
"from": "human",
"value": "人类指令"
},
{
"from": "gpt",
"value": "模型回答"
},
{
"from": "human",
"value": "人类指令"
}
],
"chosen": {
"from": "gpt",
"value": "优质回答"
},
"rejected": {
"from": "gpt",
"value": "劣质回答"
}
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"formatting": "sharegpt",
"ranking": true,
"columns": {
"messages": "conversations",
"chosen": "chosen",
"rejected": "rejected"
}
}
```
### OpenAI 格式
OpenAI 格式仅仅是 sharegpt 格式的一种特殊情况，其中第一条消息可能是系统提示词。
```json
[
{
"messages": [
{
"role": "system",
"content": "系统提示词（选填）"
},
{
"role": "user",
"content": "人类指令"
},
{
"role": "assistant",
"content": "模型回答"
}
]
}
]
```
对于上述格式的数据，`dataset_info.json` 中的*数据集描述*应为：
```json
"数据集名称": {
"file_name": "data.json",
"formatting": "sharegpt",
"columns": {
"messages": "messages"
},
"tags": {
"role_tag": "role",
"content_tag": "content",
"user_tag": "user",
"assistant_tag": "assistant",
"system_tag": "system"
}
}
```
Sharegpt 格式中的 KTO 数据集和多模态数据集与 alpaca 格式的类似。
预训练数据集**不支持** sharegpt 格式。
pretrain/LLaMA-Factory/data/c4_demo.json 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
pretrain/LLaMA-Factory/data/dataset_info.json 0 → 100644
{
"c4_demo": {
"file_name": "c4_demo.json",
"columns": {
"prompt": "text"
}
}
}
pretrain/LLaMA-Factory/examples/accelerate/fsdp_config.yaml deleted 100644 → 0
compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_backward_prefetch: BACKWARD_PRE
fsdp_forward_prefetch: false
fsdp_cpu_ram_efficient_loading: true
fsdp_offload_params: true # offload may affect training speed
fsdp_sharding_strategy: FULL_SHARD
fsdp_state_dict_type: FULL_STATE_DICT
fsdp_sync_module_states: true
fsdp_use_orig_params: true
machine_rank: 0
main_training_function: main
mixed_precision: fp16 # or bf16
num_machines: 1 # the number of nodes
num_processes: 2 # the number of GPUs in all nodes
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
pretrain/LLaMA-Factory/examples/deepspeed/ds_z0_config.json deleted 100644 → 0
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 0,
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"contiguous_gradients": true,
"round_robin_gradients": true
}
}
\ No newline at end of file
pretrain/LLaMA-Factory/examples/deepspeed/ds_z2_config.json deleted 100644 → 0
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"contiguous_gradients": true,
"round_robin_gradients": true
}
}
\ No newline at end of file
pretrain/LLaMA-Factory/examples/deepspeed/ds_z2_offload_config.json deleted 100644 → 0
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"contiguous_gradients": true,
"round_robin_gradients": true
}
}
\ No newline at end of file
pretrain/LLaMA-Factory/examples/deepspeed/ds_z3_config.json deleted 100644 → 0
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 3,
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
}
}
\ No newline at end of file
pretrain/LLaMA-Factory/examples/deepspeed/ds_z3_offload_config.json deleted 100644 → 0
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
}
}
\ No newline at end of file
pretrain/LLaMA-Factory/examples/extras/adam_mini/qwen2_full_sft.yaml deleted 100644 → 0
### model
model_name_or_path: Qwen/Qwen2-1.5B-Instruct
### method
stage: sft
do_train: true
finetuning_type: full
use_adam_mini: true
### dataset
dataset: identity,alpaca_en_demo
template: qwen
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/qwen2-1_5b/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/badam/llama3_full_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: full
use_badam: true
badam_mode: layer
badam_switch_mode: ascending
badam_switch_interval: 50
badam_verbose: 2
# deepspeed: examples/deepspeed/ds_z3_config.json
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/fsdp_qlora/llama3_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
quantization_bit: 4
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/fsdp_qlora/train.sh deleted 100644 → 0
#!/bin/bash
# DO NOT use GPTQ/AWQ model in FSDP+QLoRA
CUDA_VISIBLE_DEVICES=0,1 accelerate launch \
--config_file examples/accelerate/fsdp_config.yaml \
src/train.py examples/extras/fsdp_qlora/llama3_lora_sft.yaml
pretrain/LLaMA-Factory/examples/extras/galore/llama3_full_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: full
use_galore: true
galore_layerwise: true
galore_target: mlp,self_attn
galore_rank: 128
galore_scale: 2.0
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 1
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
pure_bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/llama_pro/expand.sh deleted 100644 → 0
#!/bin/bash
python scripts/llama_pro.py \
--model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
--output_dir models/llama3-8b-pro \
--num_expand 8
pretrain/LLaMA-Factory/examples/extras/llama_pro/llama3_freeze_sft.yaml deleted 100644 → 0
### model
model_name_or_path: models/llama3-8b-pro
### method
stage: sft
do_train: true
finetuning_type: freeze
freeze_trainable_layers: 8
freeze_trainable_modules: all
use_llama_pro: true
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b-pro/freeze/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/loraplus/llama3_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
loraplus_lr_ratio: 16.0
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/mod/llama3_full_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: full
mixture_of_depths: convert
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b-mod/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
optim: paged_adamw_8bit
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
pure_bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/extras/pissa/init.sh deleted 100644 → 0
#!/bin/bash
python scripts/pissa_init.py \
--model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
--output_dir models/llama3-8b-pissa
pretrain/LLaMA-Factory/examples/extras/pissa/llama3_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
pissa_init: true
pissa_iter: 16
pissa_convert: true
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/inference/llama3.yaml deleted 100644 → 0
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
template: llama3
pretrain/LLaMA-Factory/examples/inference/llama3_lora_sft.yaml deleted 100644 → 0
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
template: llama3
finetuning_type: lora
pretrain/LLaMA-Factory/examples/inference/llama3_vllm.yaml deleted 100644 → 0
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
template: llama3
infer_backend: vllm
vllm_enforce_eager: true
pretrain/LLaMA-Factory/examples/inference/llava1_5.yaml deleted 100644 → 0
model_name_or_path: llava-hf/llava-1.5-7b-hf
template: vicuna
visual_inputs: true
pretrain/LLaMA-Factory/examples/merge_lora/llama3_gptq.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
template: llama3
### export
export_dir: models/llama3_gptq
export_quantization_bit: 4
export_quantization_dataset: data/c4_demo.json
export_size: 2
export_device: cpu
export_legacy_format: false
pretrain/LLaMA-Factory/examples/merge_lora/llama3_lora_sft.yaml deleted 100644 → 0
### Note: DO NOT use quantized model or quantization_bit when merging lora adapters
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
template: llama3
finetuning_type: lora
### export
export_dir: models/llama3_lora_sft
export_size: 2
export_device: cpu
export_legacy_format: false
pretrain/LLaMA-Factory/examples/train_full/llama3_full_predict.yaml deleted 100644 → 0
### model
model_name_or_path: saves/llama3-8b/full/sft
### method
stage: sft
do_predict: true
finetuning_type: full
### dataset
eval_dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 50
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/full/predict
overwrite_output_dir: true
### eval
per_device_eval_batch_size: 1
predict_with_generate: true
pretrain/LLaMA-Factory/examples/train_full/llama3_full_sft_ds3.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_dpo.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: dpo
do_train: true
finetuning_type: lora
lora_target: all
pref_beta: 0.1
pref_loss: sigmoid # choices: [sigmoid (dpo), orpo, simpo]
### dataset
dataset: dpo_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/dpo
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 5.0e-6
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_eval.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
### method
finetuning_type: lora
### dataset
task: mmlu_test # choices: [mmlu_test, ceval_validation, cmmlu_test]
template: fewshot
lang: en
n_shot: 5
### output
save_dir: saves/llama3-8b/lora/eval
### eval
batch_size: 4
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_kto.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: kto
do_train: true
finetuning_type: lora
lora_target: all
pref_beta: 0.1
### dataset
dataset: kto_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/kto
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 5.0e-6
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_ppo.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
reward_model: saves/llama3-8b/lora/reward
### method
stage: ppo
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/ppo
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### generate
max_new_tokens: 512
top_k: 0
top_p: 0.9
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_predict.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
### method
stage: sft
do_predict: true
finetuning_type: lora
### dataset
eval_dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 50
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/predict
overwrite_output_dir: true
### eval
per_device_eval_batch_size: 1
predict_with_generate: true
ddp_timeout: 180000000
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_pretrain.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: pt
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: c4_demo
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/pretrain
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_reward.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: rm
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: dpo_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/reward
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_sft_ds0.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
deepspeed: examples/deepspeed/ds_z0_config.json
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_lora_sft_ds3.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
deepspeed: examples/deepspeed/ds_z3_config.json
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/llama3_preprocess.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
tokenized_path: saves/llama3-8b/dataset/sft
### output
output_dir: saves/llama3-8b/lora/sft
overwrite_output_dir: true
pretrain/LLaMA-Factory/examples/train_lora/llava1_5_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: llava-hf/llava-1.5-7b-hf
visual_inputs: true
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: mllm_demo
template: vicuna
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llava1_5-7b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_lora/qwen2vl_lora_sft.yaml deleted 100644 → 0
### model
model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
visual_inputs: true
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: mllm_demo
template: qwen2_vl
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/qwen2_vl-7b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_qlora/llama3_lora_sft_aqlm.yaml deleted 100644 → 0
### model
model_name_or_path: ISTA-DASLab/Meta-Llama-3-8B-Instruct-AQLM-2Bit-1x16
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_qlora/llama3_lora_sft_awq.yaml deleted 100644 → 0
### model
model_name_or_path: TechxGenus/Meta-Llama-3-8B-Instruct-AWQ
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_qlora/llama3_lora_sft_gptq.yaml deleted 100644 → 0
### model
model_name_or_path: TechxGenus/Meta-Llama-3-8B-Instruct-GPTQ
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/examples/train_qlora/llama3_lora_sft_otfq.yaml deleted 100644 → 0
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
quantization_bit: 4
quantization_method: bitsandbytes # choices: [bitsandbytes (4/8), hqq (2/3/4/5/6/8), eetq (8)]
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
pretrain/LLaMA-Factory/pretrain.slurm 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p r8nv-gpu-dedicated # Submit to 'nv-gpu' Partitiion
#SBATCH -t 7-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=4 # Request N nodes
#SBATCH --gres=gpu:8 # Request M GPU per node
#SBATCH --ntasks=4 #
### #SBATCH --ntasks-per-node=1 # Request P tasks per node
#SBATCH --cpus-per-task=16 # Request Q core per task; means that P*Q cores per node
#SBATCH --qos=normal # Request QOS Type
### #SBATCH --constraint="IB&A100"
#SBATCH --nodelist=r8a100-b[00,02,05-06]
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=r8a100-b[00,02,05-06]
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
# set constraint for RTX8000 to meet my cuda
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
##- language
module load python3/3.8.16
##- CUDA
module load cuda-cudnn/11.6-8.4.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
CUDA_VISIBLE_DEVICES = 8
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
# [EDIT HERE(TODO)]
# export http_proxy=10.200.22.22:18421
# export https_proxy=10.200.22.22:18421
export GIT_SSL_NO_VERIFY=1
export HF_ENDPOINT=https://hf-mirror.com
master_node=$(scontrol show hostname $SLURM_NODELIST | head -n 1)
# Resolve the hostname to an IP address
master_ip=$(getent ahosts $master_node | grep "^[0-9]" | head -n 1 | awk '{ print $1 }')
export TASK_NAME=minicpm_pretrain
echo $master_node
echo $master_ip
scontrol show hostnames $SLURM_NODELIST
srun bash -c "FORCE_TORCHRUN=1 NNODES=4 RANK=\$SLURM_NODEID MASTER_ADDR=$master_ip MASTER_PORT=29500 llamafactory-cli train configs/$TASK_NAME.yaml > ret_one/${TASK_NAME}_log.\$SLURM_NODEID 2>&1"
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
pretrain/LLaMA-Factory/src/llamafactory/data/__init__.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .collator import (
CustomDataCollatorForSeq2Seq,
KTODataCollatorWithPadding,
PairwiseDataCollatorWithPadding,
SFTDataCollatorWith4DAttentionMask,
)
from .data_utils import Role, split_dataset
from .loader import get_dataset
from .template import TEMPLATES, Template, get_template_and_fix_tokenizer
__all__ = [
"CustomDataCollatorForSeq2Seq",
"KTODataCollatorWithPadding",
"PairwiseDataCollatorWithPadding",
"SFTDataCollatorWith4DAttentionMask",
"Role",
"split_dataset",
"get_dataset",
"TEMPLATES",
"Template",
"get_template_and_fix_tokenizer",
]
pretrain/LLaMA-Factory/src/llamafactory/data/collator.py 0 → 100644
# Copyright 2024 OpenAccess AI Collective and the LlamaFactory team.
#
# This code is inspired by the OpenAccess AI Collective's axolotl library.
# https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Any, Dict, Literal, Sequence
import torch
from transformers import DataCollatorForSeq2Seq
def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype: "torch.dtype") -> "torch.Tensor":
r"""
Expands the attention mask with indices from (batch_size, seq_len) to (batch_size, 1, seq_len, seq_len),
while handles packed sequences and transforms the mask to lower triangular form to prevent future peeking.
e.g.
```python
# input
[[1, 1, 2, 2, 2, 0]]
# output
[
[
[
[o, x, x, x, x, x],
[o, o, x, x, x, x],
[x, x, o, x, x, x],
[x, x, o, o, x, x],
[x, x, o, o, o, x],
[x, x, x, x, x, x],
]
]
]
```
where `o` equals to `0.0`, `x` equals to `min_dtype`.
"""
bsz, seq_len = attention_mask_with_indices.size()
min_dtype = torch.finfo(dtype).min
expanded_mask = attention_mask_with_indices[:, None, None, :].expand(bsz, 1, seq_len, seq_len)
# Create a binary mask from the original mask where zeros remain zeros and all other values are set to one
padding_mask = torch.where(expanded_mask != 0, 1, 0)
# Create a block-diagonal mask.
attention_mask_4d = torch.eq(expanded_mask, expanded_mask.transpose(-1, -2)).int() * padding_mask
# Use the lower triangular mask to zero out the upper triangular part
attention_mask_4d *= torch.tril(torch.ones((seq_len, seq_len), dtype=torch.long))
# Invert the attention mask.
attention_mask_4d = torch.where(attention_mask_4d != 0, torch.tensor(0, dtype=dtype), min_dtype)
return attention_mask_4d
@dataclass
class CustomDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
r"""
Data collator for custom models (like Qwen2-VL).
"""
def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
image_grid_thw = None # TODO: better handle various VLMs
if "image_grid_thw" in features[0]:
image_grid_thw_list = [
torch.Tensor(feature["image_grid_thw"]).long()
for feature in features
if feature["image_grid_thw"][0][0] > 0
]
pixel_values_list = [
torch.Tensor(feature["pixel_values"]) for feature in features if feature["image_grid_thw"][0][0] > 0
]
if image_grid_thw_list:
image_grid_thw = torch.cat(image_grid_thw_list, dim=0)
pixel_values = torch.cat(pixel_values_list, dim=0)
else:
image_grid_thw = None
pixel_values = None
features = [
{key: feature[key] for key in feature if key not in ["image_grid_thw", "pixel_values"]}
for feature in features
]
features = super().__call__(features)
if image_grid_thw is not None:
features["image_grid_thw"] = image_grid_thw
features["pixel_values"] = pixel_values
return features
@dataclass
class SFTDataCollatorWith4DAttentionMask(CustomDataCollatorForSeq2Seq):
r"""
Data collator for 4d attention mask.
"""
block_diag_attn: bool = False
attn_implementation: Literal["eager", "sdpa", "flash_attention_2"] = "eager"
compute_dtype: "torch.dtype" = torch.float32
def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
features = super().__call__(features)
if self.block_diag_attn and self.attn_implementation != "flash_attention_2":
features["attention_mask"] = prepare_4d_attention_mask(features["attention_mask"], self.compute_dtype)
return features
@dataclass
class PairwiseDataCollatorWithPadding(CustomDataCollatorForSeq2Seq):
r"""
Data collator for pairwise data.
"""
def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
r"""
Pads batched data to the longest sequence in the batch.
We generate 2 * n examples where the first n examples represent chosen examples and
the last n examples represent rejected examples.
"""
concatenated_features = []
for key in ("chosen", "rejected"):
for feature in features:
target_feature = {
"input_ids": feature["{}_input_ids".format(key)],
"attention_mask": feature["{}_attention_mask".format(key)],
"labels": feature["{}_labels".format(key)],
}
if "pixel_values" in feature: # image data are same for chosen and rejected
target_feature["pixel_values"] = feature["pixel_values"]
if "image_grid_thw" in feature:
target_feature["image_grid_thw"] = feature["image_grid_thw"]
if "{}_token_type_ids".format(key) in feature:
target_feature["token_type_ids"] = feature["{}_token_type_ids".format(key)]
concatenated_features.append(target_feature)
return super().__call__(concatenated_features)
@dataclass
class KTODataCollatorWithPadding(CustomDataCollatorForSeq2Seq):
r"""
Data collator for KTO data.
"""
def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, "torch.Tensor"]:
target_features = []
kl_features = []
kto_tags = []
for feature in features:
target_feature = {
"input_ids": feature["input_ids"],
"attention_mask": feature["attention_mask"],
"labels": feature["labels"],
}
kl_feature = {
"input_ids": feature["kl_input_ids"],
"attention_mask": feature["kl_attention_mask"],
"labels": feature["kl_labels"],
}
if "pixel_values" in feature:
target_feature["pixel_values"] = feature["pixel_values"]
if "image_grid_thw" in feature:
target_feature["image_grid_thw"] = feature["image_grid_thw"]
if "token_type_ids" in feature:
target_feature["token_type_ids"] = feature["token_type_ids"]
kl_feature["token_type_ids"] = feature["kl_token_type_ids"]
target_features.append(target_feature)
kl_features.append(kl_feature)
kto_tags.append(feature["kto_tags"])
batch = super().__call__(target_features)
kl_batch = super().__call__(kl_features)
batch["kl_input_ids"] = kl_batch["input_ids"]
batch["kl_attention_mask"] = kl_batch["attention_mask"]
batch["kl_labels"] = kl_batch["labels"]
if "token_type_ids" in batch:
batch["kl_token_type_ids"] = kl_batch["token_type_ids"]
batch["kto_tags"] = torch.tensor(kto_tags)
return batch
pretrain/LLaMA-Factory/src/llamafactory/data/data_utils.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from enum import Enum, unique
from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Set, TypedDict, Union
from datasets import DatasetDict, concatenate_datasets, interleave_datasets
from ..extras.logging import get_logger
if TYPE_CHECKING:
from datasets import Dataset, IterableDataset
from ..hparams import DataArguments
logger = get_logger(__name__)
SLOTS = Sequence[Union[str, Set[str], Dict[str, str]]]
@unique
class Role(str, Enum):
USER = "user"
ASSISTANT = "assistant"
SYSTEM = "system"
FUNCTION = "function"
OBSERVATION = "observation"
class DatasetModule(TypedDict):
train_dataset: Optional[Union["Dataset", "IterableDataset"]]
eval_dataset: Optional[Union["Dataset", "IterableDataset"]]
def merge_dataset(
all_datasets: List[Union["Dataset", "IterableDataset"]], data_args: "DataArguments", seed: int
) -> Union["Dataset", "IterableDataset"]:
if len(all_datasets) == 1:
return all_datasets[0]
elif data_args.mix_strategy == "concat":
if data_args.streaming:
logger.warning("The samples between different datasets will not be mixed in streaming mode.")
return concatenate_datasets(all_datasets)
elif data_args.mix_strategy.startswith("interleave"):
if not data_args.streaming:
logger.warning("We recommend using `mix_strategy=concat` in non-streaming mode.")
return interleave_datasets(
datasets=all_datasets,
probabilities=data_args.interleave_probs,
seed=seed,
stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
)
else:
raise ValueError("Unknown mixing strategy.")
def split_dataset(
dataset: Union["Dataset", "IterableDataset"], data_args: "DataArguments", seed: int
) -> "DatasetDict":
r"""
Splits the dataset and returns a dataset dict containing train set (required) and validation set (optional).
"""
if data_args.streaming:
dataset = dataset.shuffle(buffer_size=data_args.buffer_size, seed=seed)
val_set = dataset.take(int(data_args.val_size))
train_set = dataset.skip(int(data_args.val_size))
return DatasetDict({"train": train_set, "validation": val_set})
else:
val_size = int(data_args.val_size) if data_args.val_size > 1 else data_args.val_size
dataset = dataset.train_test_split(test_size=val_size, seed=seed)
return DatasetDict({"train": dataset["train"], "validation": dataset["test"]})
pretrain/LLaMA-Factory/src/llamafactory/data/formatter.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import re
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import List, Literal, Optional, Tuple, Union
from .data_utils import SLOTS
from .tool_utils import DefaultToolUtils, GLM4ToolUtils
@dataclass
class Formatter(ABC):
slots: SLOTS = field(default_factory=list)
tool_format: Optional[Literal["default", "glm4"]] = None
@abstractmethod
def apply(self, **kwargs) -> SLOTS: ...
def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
raise NotImplementedError
@dataclass
class EmptyFormatter(Formatter):
def __post_init__(self):
has_placeholder = False
for slot in filter(lambda s: isinstance(s, str), self.slots):
if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
has_placeholder = True
if has_placeholder:
raise ValueError("Empty formatter should not contain any placeholder.")
def apply(self, **kwargs) -> SLOTS:
return self.slots
@dataclass
class StringFormatter(Formatter):
def __post_init__(self):
has_placeholder = False
for slot in filter(lambda s: isinstance(s, str), self.slots):
if re.search(r"\{\{[a-zA-Z_][a-zA-Z0-9_]*\}\}", slot):
has_placeholder = True
if not has_placeholder:
raise ValueError("A placeholder is required in the string formatter.")
def apply(self, **kwargs) -> SLOTS:
elements = []
for slot in self.slots:
if isinstance(slot, str):
for name, value in kwargs.items():
if not isinstance(value, str):
raise RuntimeError("Expected a string, got {}".format(value))
slot = slot.replace("{{" + name + "}}", value, 1)
elements.append(slot)
elif isinstance(slot, (dict, set)):
elements.append(slot)
else:
raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
return elements
@dataclass
class FunctionFormatter(Formatter):
def __post_init__(self):
if self.tool_format == "default":
self.slots = DefaultToolUtils.get_function_slots() + self.slots
elif self.tool_format == "glm4":
self.slots = GLM4ToolUtils.get_function_slots() + self.slots
else:
raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
def apply(self, **kwargs) -> SLOTS:
content = kwargs.pop("content")
functions: List[Tuple[str, str]] = []
try:
tool_calls = json.loads(content)
if not isinstance(tool_calls, list): # parallel function call
tool_calls = [tool_calls]
for tool_call in tool_calls:
functions.append((tool_call["name"], json.dumps(tool_call["arguments"], ensure_ascii=False)))
except json.JSONDecodeError:
functions = []
elements = []
for name, arguments in functions:
for slot in self.slots:
if isinstance(slot, str):
slot = slot.replace("{{name}}", name).replace("{{arguments}}", arguments)
elements.append(slot)
elif isinstance(slot, (dict, set)):
elements.append(slot)
else:
raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
return elements
@dataclass
class ToolFormatter(Formatter):
def __post_init__(self):
if self.tool_format == "default":
self._tool_formatter = DefaultToolUtils.tool_formatter
self._tool_extractor = DefaultToolUtils.tool_extractor
elif self.tool_format == "glm4":
self._tool_formatter = GLM4ToolUtils.tool_formatter
self._tool_extractor = GLM4ToolUtils.tool_extractor
else:
raise NotImplementedError("Tool format {} was not found.".format(self.tool_format))
def apply(self, **kwargs) -> SLOTS:
content = kwargs.pop("content")
try:
tools = json.loads(content)
return [self._tool_formatter(tools) if len(tools) != 0 else ""]
except json.JSONDecodeError:
return [""]
def extract(self, content: str) -> Union[str, List[Tuple[str, str]]]:
return self._tool_extractor(content)
pretrain/LLaMA-Factory/src/llamafactory/data/mm_plugin.py 0 → 100644
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
from PIL.Image import Image
from transformers import ProcessorMixin
from ..extras.constants import IGNORE_INDEX, IMAGE_PLACEHOLDER
from ..extras.packages import is_pillow_available
if is_pillow_available():
import torch
from PIL import Image
if TYPE_CHECKING:
from PIL.Image import Image as ImageObject
from transformers import PreTrainedTokenizer, ProcessorMixin
from transformers.image_processing_utils import BaseImageProcessor
def get_pixel_values(images: Sequence["ImageObject"], processor: "ProcessorMixin") -> "torch.Tensor":
r"""
Processes visual inputs. (currently only supports a single image)
Returns:
pixel_values: tensor with shape (B, C, H, W)
"""
image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
image = images[0] if len(images) != 0 else Image.new("RGB", (100, 100), (255, 255, 255))
return image_processor([image], return_tensors="pt")["pixel_values"]
def get_paligemma_token_type_ids(input_len: int, processor: "ProcessorMixin") -> List[List[int]]:
r"""
Gets paligemma token type ids for computing loss.
Returns:
token_type_ids: shape (1, seq_len)
"""
image_seq_length = getattr(processor, "image_seq_length")
return [[0] * image_seq_length + [1] * (input_len - image_seq_length)]
def get_qwen2vl_image_inputs(
images: Sequence["ImageObject"], processor: "ProcessorMixin"
) -> Dict[str, "torch.Tensor"]:
r"""
Processes qwen2-vl visual inputs. Supports multiple images.
Returns:
pixel_values: tensor with shape (num_patches, patch_dim)
image_grid_thw: tensot with shape (num_images, 3), where the three numbers are time, width, height
It holds num_patches == torch.prod(image_grid_thw)
"""
image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
if len(images) != 0:
image_inputs = image_processor(images=images, return_tensors="pt")
else:
image = Image.new("RGB", (56, 56), (255, 255, 255))
image_inputs = image_processor(images=[image], return_tensors="pt")
image_inputs["image_grid_thw"][0][0] = 0 # fake image
return {"pixel_values": image_inputs["pixel_values"], "image_grid_thw": image_inputs["image_grid_thw"]}
class BasePlugin:
def __init__(self, image_token: str) -> None:
self.image_token = image_token
def process_messages(
self,
messages: Sequence[Dict[str, str]],
images: Sequence["ImageObject"],
processor: Optional["ProcessorMixin"],
) -> List[Dict[str, str]]:
r"""
Pre-processes input messages before tokenization for VLMs.
"""
return messages
def process_token_ids(
self,
input_ids: List[int],
labels: Optional[List[int]],
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
) -> Tuple[List[int], Optional[List[int]]]:
r"""
Pre-processes token ids after tokenization for VLMs.
"""
return input_ids, labels
def get_mm_inputs(
self,
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> Dict[str, Any]:
r"""
Builds batched multimodal inputs for VLMs.
"""
return {}
def process_model_inputs(
self,
model_inputs: Dict[str, List[Any]],
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> None:
r"""
Appends multimodal inputs to model inputs for VLMs.
"""
return
class LlavaPlugin(BasePlugin):
def process_messages(
self,
messages: Sequence[Dict[str, str]],
images: Sequence["ImageObject"],
processor: Optional["ProcessorMixin"],
) -> List[Dict[str, str]]:
image_count = 0
new_messages = []
for message in messages:
content = message["content"]
while IMAGE_PLACEHOLDER in content:
image_count += 1
if image_count > 1:
raise ValueError("Llava model only accepts one image per sample.")
content = content.replace(IMAGE_PLACEHOLDER, self.image_token, 1)
new_messages.append({"role": message["role"], "content": content})
return new_messages
def get_mm_inputs(
self,
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> Dict[str, Any]:
return {"pixel_values": get_pixel_values(images, processor)}
def process_model_inputs(
self,
model_inputs: Dict[str, List[Any]],
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> None:
mm_inputs = self.get_mm_inputs(images, feature_seqlens, processor)
model_inputs["pixel_values"].append(mm_inputs["pixel_values"][0])
class PaliGemmaPlugin(BasePlugin):
def process_messages(
self,
messages: Sequence[Dict[str, str]],
images: Sequence["ImageObject"],
processor: Optional["ProcessorMixin"],
) -> List[Dict[str, str]]:
image_count = 0
new_messages = []
for message in messages:
content = message["content"]
while IMAGE_PLACEHOLDER in content:
image_count += 1
if image_count > 1:
raise ValueError("PaliGemma model only accepts one image per sample.")
content = content.replace(IMAGE_PLACEHOLDER, "", 1)
new_messages.append({"role": message["role"], "content": content})
return new_messages
def process_token_ids(
self,
input_ids: List[int],
labels: Optional[List[int]],
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
) -> Tuple[List[int], Optional[List[int]]]:
image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
image_seq_length: int = getattr(image_processor, "image_seq_length")
image_token_id = tokenizer.convert_tokens_to_ids(self.image_token)
input_ids = [image_token_id] * image_seq_length + input_ids
if labels is not None:
labels = [IGNORE_INDEX] * image_seq_length + labels
return input_ids, labels
def get_mm_inputs(
self,
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> Dict[str, Any]:
mm_inputs = {"pixel_values": get_pixel_values(images, processor)}
for feature_name, feature_length in feature_seqlens.items():
mm_inputs[feature_name] = get_paligemma_token_type_ids(feature_length, processor)
return mm_inputs
def process_model_inputs(
self,
model_inputs: Dict[str, List[Any]],
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> None:
mm_inputs = self.get_mm_inputs(images, feature_seqlens, processor)
model_inputs["pixel_values"].append(mm_inputs["pixel_values"][0])
for feature_name in feature_seqlens.keys():
model_inputs[feature_name].append(mm_inputs[feature_name][0])
class Qwen2vlPlugin(BasePlugin):
def process_messages(
self,
messages: Sequence[Dict[str, str]],
images: Sequence["ImageObject"],
processor: Optional["ProcessorMixin"],
) -> List[Dict[str, str]]:
image_processor: "BaseImageProcessor" = getattr(processor, "image_processor")
merge_length: int = getattr(image_processor, "merge_size") ** 2
if len(images) > 0:
image_grid_thw = get_qwen2vl_image_inputs(images, processor)["image_grid_thw"]
index = 0
new_messages = []
for message in messages:
content = message["content"]
while IMAGE_PLACEHOLDER in content:
content = content.replace(
IMAGE_PLACEHOLDER,
"<|vision_start|>{}<|vision_end|>".format(
self.image_token * (image_grid_thw[index].prod() // merge_length)
),
1,
)
index += 1
new_messages.append({"role": message["role"], "content": content})
return new_messages
def get_mm_inputs(
self,
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> Dict[str, Any]:
return get_qwen2vl_image_inputs(images, processor)
def process_model_inputs(
self,
model_inputs: Dict[str, List[Any]],
images: Sequence["ImageObject"],
feature_seqlens: Dict[str, int],
processor: Optional["ProcessorMixin"],
) -> None:
mm_inputs = self.get_mm_inputs(images, feature_seqlens, processor)
model_inputs["pixel_values"].append(mm_inputs["pixel_values"])
model_inputs["image_grid_thw"].append(mm_inputs["image_grid_thw"])
PLUGINS = {
"llava": LlavaPlugin,
"paligemma": PaliGemmaPlugin,
"qwen2_vl": Qwen2vlPlugin,
}
def get_mm_plugin(name: str, image_token: str) -> "BasePlugin":
if name not in PLUGINS:
raise ValueError("{} not found.".format(name))
return PLUGINS[name](image_token)
pretrain/LLaMA-Factory/src/llamafactory/data/parser.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import os
from dataclasses import dataclass
from typing import Any, Dict, List, Literal, Optional, Sequence
from transformers.utils import cached_file
from ..extras.constants import DATA_CONFIG
from ..extras.misc import use_modelscope
@dataclass
class DatasetAttr:
r"""
Dataset attributes.
"""
# basic configs
load_from: Literal["hf_hub", "ms_hub", "script", "file"]
dataset_name: str
formatting: Literal["alpaca", "sharegpt"] = "alpaca"
ranking: bool = False
# extra configs
subset: Optional[str] = None
split: str = "train"
folder: Optional[str] = None
num_samples: Optional[int] = None
# common columns
system: Optional[str] = None
tools: Optional[str] = None
images: Optional[str] = None
# rlhf columns
chosen: Optional[str] = None
rejected: Optional[str] = None
kto_tag: Optional[str] = None
# alpaca columns
prompt: Optional[str] = "instruction"
query: Optional[str] = "input"
response: Optional[str] = "output"
history: Optional[str] = None
# sharegpt columns
messages: Optional[str] = "conversations"
# sharegpt tags
role_tag: Optional[str] = "from"
content_tag: Optional[str] = "value"
user_tag: Optional[str] = "human"
assistant_tag: Optional[str] = "gpt"
observation_tag: Optional[str] = "observation"
function_tag: Optional[str] = "function_call"
system_tag: Optional[str] = "system"
def __repr__(self) -> str:
return self.dataset_name
def set_attr(self, key: str, obj: Dict[str, Any], default: Optional[Any] = None) -> None:
setattr(self, key, obj.get(key, default))
def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -> List["DatasetAttr"]:
r"""
Gets the attributes of the datasets.
"""
if dataset_names is None:
dataset_names = []
if dataset_dir == "ONLINE":
dataset_info = None
else:
if dataset_dir.startswith("REMOTE:"):
config_path = cached_file(path_or_repo_id=dataset_dir[7:], filename=DATA_CONFIG, repo_type="dataset")
else:
config_path = os.path.join(dataset_dir, DATA_CONFIG)
try:
with open(config_path, "r") as f:
dataset_info = json.load(f)
except Exception as err:
if len(dataset_names) != 0:
raise ValueError("Cannot open {} due to {}.".format(config_path, str(err)))
dataset_info = None
dataset_list: List["DatasetAttr"] = []
for name in dataset_names:
if dataset_info is None: # dataset_dir is ONLINE
load_from = "ms_hub" if use_modelscope() else "hf_hub"
dataset_attr = DatasetAttr(load_from, dataset_name=name)
dataset_list.append(dataset_attr)
continue
if name not in dataset_info:
raise ValueError("Undefined dataset {} in {}.".format(name, DATA_CONFIG))
has_hf_url = "hf_hub_url" in dataset_info[name]
has_ms_url = "ms_hub_url" in dataset_info[name]
if has_hf_url or has_ms_url:
if (use_modelscope() and has_ms_url) or (not has_hf_url):
dataset_attr = DatasetAttr("ms_hub", dataset_name=dataset_info[name]["ms_hub_url"])
else:
dataset_attr = DatasetAttr("hf_hub", dataset_name=dataset_info[name]["hf_hub_url"])
elif "script_url" in dataset_info[name]:
dataset_attr = DatasetAttr("script", dataset_name=dataset_info[name]["script_url"])
else:
dataset_attr = DatasetAttr("file", dataset_name=dataset_info[name]["file_name"])
dataset_attr.set_attr("formatting", dataset_info[name], default="alpaca")
dataset_attr.set_attr("ranking", dataset_info[name], default=False)
dataset_attr.set_attr("subset", dataset_info[name])
dataset_attr.set_attr("split", dataset_info[name], default="train")
dataset_attr.set_attr("folder", dataset_info[name])
dataset_attr.set_attr("num_samples", dataset_info[name])
if "columns" in dataset_info[name]:
column_names = ["system", "tools", "images", "chosen", "rejected", "kto_tag"]
if dataset_attr.formatting == "alpaca":
column_names.extend(["prompt", "query", "response", "history"])
else:
column_names.extend(["messages"])
for column_name in column_names:
dataset_attr.set_attr(column_name, dataset_info[name]["columns"])
if dataset_attr.formatting == "sharegpt" and "tags" in dataset_info[name]:
tag_names = (
"role_tag",
"content_tag",
"user_tag",
"assistant_tag",
"observation_tag",
"function_tag",
"system_tag",
)
for tag in tag_names:
dataset_attr.set_attr(tag, dataset_info[name]["tags"])
dataset_list.append(dataset_attr)
return dataset_list
pretrain/LLaMA-Factory/src/llamafactory/data/preprocess.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from functools import partial
from typing import TYPE_CHECKING, Callable, Literal, Optional, Tuple
from .processors.feedback import preprocess_feedback_dataset
from .processors.pairwise import preprocess_pairwise_dataset, print_pairwise_dataset_example
from .processors.pretrain import preprocess_pretrain_dataset
from .processors.supervised import (
preprocess_packed_supervised_dataset,
preprocess_supervised_dataset,
print_supervised_dataset_example,
)
from .processors.unsupervised import preprocess_unsupervised_dataset, print_unsupervised_dataset_example
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ..hparams import DataArguments
from .template import Template
def get_preprocess_and_print_func(
data_args: "DataArguments",
stage: Literal["pt", "sft", "rm", "ppo", "kto"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
do_generate: bool = False,
) -> Tuple[Callable, Callable]:
if stage == "pt":
preprocess_func = partial(
preprocess_pretrain_dataset,
tokenizer=tokenizer,
data_args=data_args,
)
print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
elif stage == "sft" and not do_generate:
if data_args.packing:
if data_args.neat_packing:
from datasets.arrow_writer import OptimizedTypedSequence, TypedSequence
def __init__(self, data, **kwargs):
return TypedSequence.__init__(
self,
data,
type=kwargs.pop("type", None),
try_type=kwargs.pop("try_type", None),
optimized_int_type=kwargs.pop("optimized_int_type", None),
)
OptimizedTypedSequence.__init__ = __init__
preprocess_func = partial(
preprocess_packed_supervised_dataset,
template=template,
tokenizer=tokenizer,
data_args=data_args,
)
else:
preprocess_func = partial(
preprocess_supervised_dataset,
template=template,
tokenizer=tokenizer,
processor=processor,
data_args=data_args,
)
print_function = partial(print_supervised_dataset_example, tokenizer=tokenizer)
elif stage == "rm":
preprocess_func = partial(
preprocess_pairwise_dataset,
template=template,
tokenizer=tokenizer,
processor=processor,
data_args=data_args,
)
print_function = partial(print_pairwise_dataset_example, tokenizer=tokenizer)
elif stage == "kto":
preprocess_func = partial(
preprocess_feedback_dataset,
template=template,
tokenizer=tokenizer,
processor=processor,
data_args=data_args,
)
print_function = partial(print_supervised_dataset_example, tokenizer=tokenizer)
else:
preprocess_func = partial(
preprocess_unsupervised_dataset,
template=template,
tokenizer=tokenizer,
processor=processor,
data_args=data_args,
)
print_function = partial(print_unsupervised_dataset_example, tokenizer=tokenizer)
return preprocess_func, print_function
pretrain/LLaMA-Factory/src/llamafactory/data/processors/feedback.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..template import Template
logger = get_logger(__name__)
def _encode_feedback_example(
prompt: Sequence[Dict[str, str]],
response: Sequence[Dict[str, str]],
kl_response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int], bool]:
if response[0]["content"]: # desired example
kto_tag = True
messages = prompt + [response[0]]
else: # undesired example
kto_tag = False
messages = prompt + [response[1]]
if kl_response[0]["content"]:
kl_messages = prompt + [kl_response[0]]
else:
kl_messages = prompt + [kl_response[1]]
prompt_ids, response_ids = template.encode_oneturn(tokenizer, messages, system, tools)
kl_prompt_ids, kl_response_ids = template.encode_oneturn(tokenizer, kl_messages, system, tools)
if template.efficient_eos:
response_ids += [tokenizer.eos_token_id]
kl_response_ids += [tokenizer.eos_token_id]
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, tokenizer, processor)
kl_prompt_ids, _ = template.mm_plugin.process_token_ids(kl_prompt_ids, None, tokenizer, processor)
source_len, target_len = infer_seqlen(len(prompt_ids), len(response_ids), cutoff_len)
prompt_ids = prompt_ids[:source_len]
response_ids = response_ids[:target_len]
kl_source_len, kl_target_len = infer_seqlen(len(kl_prompt_ids), len(kl_response_ids), cutoff_len)
kl_prompt_ids = kl_prompt_ids[:kl_source_len]
kl_response_ids = kl_response_ids[:kl_target_len]
input_ids = prompt_ids + response_ids
labels = [IGNORE_INDEX] * source_len + response_ids
kl_input_ids = kl_prompt_ids + kl_response_ids
kl_labels = [IGNORE_INDEX] * kl_source_len + kl_response_ids
return input_ids, labels, kl_input_ids, kl_labels, kto_tag
def preprocess_feedback_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# create unrelated input-output pairs for estimating the KL term by flipping the matched pairs
kl_response = examples["response"][::-1]
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels, kl_input_ids, kl_labels, kto_tag = _encode_feedback_example(
prompt=prompt,
response=examples["response"][i],
kl_response=kl_response[i],
system=examples["system"][i],
tools=examples["tools"][i],
template=template,
tokenizer=tokenizer,
processor=processor,
cutoff_len=data_args.cutoff_len,
)
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
model_inputs["kl_input_ids"].append(kl_input_ids)
model_inputs["kl_attention_mask"].append([1] * len(kl_input_ids))
model_inputs["kl_labels"].append(kl_labels)
model_inputs["kto_tags"].append(kto_tag)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={
"token_type_ids": len(input_ids),
"kl_token_type_ids": len(kl_input_ids),
},
processor=processor,
)
desirable_num = sum([1 for tag in model_inputs["kto_tags"] if tag])
undesirable_num = len(model_inputs["kto_tags"]) - desirable_num
if desirable_num == 0 or undesirable_num == 0:
logger.warning("Your dataset only has one preference type.")
return model_inputs
pretrain/LLaMA-Factory/src/llamafactory/data/processors/pairwise.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..template import Template
logger = get_logger(__name__)
def _encode_pairwise_example(
prompt: Sequence[Dict[str, str]],
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int], List[int], List[int]]:
chosen_messages = prompt + [response[0]]
rejected_messages = prompt + [response[1]]
prompt_ids, chosen_ids = template.encode_oneturn(tokenizer, chosen_messages, system, tools)
_, rejected_ids = template.encode_oneturn(tokenizer, rejected_messages, system, tools)
if template.efficient_eos:
chosen_ids += [tokenizer.eos_token_id]
rejected_ids += [tokenizer.eos_token_id]
prompt_ids, _ = template.mm_plugin.process_token_ids(prompt_ids, None, tokenizer, processor)
# consider the response is more important
source_len, target_len = infer_seqlen(len(prompt_ids), max(len(chosen_ids), len(rejected_ids)), cutoff_len)
prompt_ids = prompt_ids[:source_len]
chosen_ids = chosen_ids[:target_len]
rejected_ids = rejected_ids[:target_len]
chosen_input_ids = prompt_ids + chosen_ids
chosen_labels = [IGNORE_INDEX] * source_len + chosen_ids
rejected_input_ids = prompt_ids + rejected_ids
rejected_labels = [IGNORE_INDEX] * source_len + rejected_ids
return chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels
def preprocess_pairwise_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# build input pairs with format `<bos> X`, `Y1 <eos>` and `Y2 <eos>`
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
chosen_input_ids, chosen_labels, rejected_input_ids, rejected_labels = _encode_pairwise_example(
prompt=prompt,
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
template=template,
tokenizer=tokenizer,
processor=processor,
cutoff_len=data_args.cutoff_len,
)
model_inputs["chosen_input_ids"].append(chosen_input_ids)
model_inputs["chosen_attention_mask"].append([1] * len(chosen_input_ids))
model_inputs["chosen_labels"].append(chosen_labels)
model_inputs["rejected_input_ids"].append(rejected_input_ids)
model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids))
model_inputs["rejected_labels"].append(rejected_labels)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={
"chosen_token_type_ids": len(chosen_input_ids),
"rejected_token_type_ids": len(rejected_input_ids),
},
processor=processor,
)
return model_inputs
def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
valid_chosen_labels = list(filter(lambda x: x != IGNORE_INDEX, example["chosen_labels"]))
valid_rejected_labels = list(filter(lambda x: x != IGNORE_INDEX, example["rejected_labels"]))
print("chosen_input_ids:\n{}".format(example["chosen_input_ids"]))
print("chosen_inputs:\n{}".format(tokenizer.decode(example["chosen_input_ids"], skip_special_tokens=False)))
print("chosen_label_ids:\n{}".format(example["chosen_labels"]))
print("chosen_labels:\n{}".format(tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)))
print("rejected_input_ids:\n{}".format(example["rejected_input_ids"]))
print("rejected_inputs:\n{}".format(tokenizer.decode(example["rejected_input_ids"], skip_special_tokens=False)))
print("rejected_label_ids:\n{}".format(example["rejected_labels"]))
print("rejected_labels:\n{}".format(tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)))
pretrain/LLaMA-Factory/src/llamafactory/data/processors/pretrain.py 0 → 100644
# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
#
# This code is inspired by the HuggingFace's transformers library.
# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from itertools import chain
from typing import TYPE_CHECKING, Any, Dict, List
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer
from ...hparams import DataArguments
def preprocess_pretrain_dataset(
examples: Dict[str, List[Any]], tokenizer: "PreTrainedTokenizer", data_args: "DataArguments"
) -> Dict[str, List[Any]]:
# build grouped texts with format `X1 X2 X3 ...` if packing is enabled
eos_token = "<|end_of_text|>" if data_args.template == "llama3" else tokenizer.eos_token
text_examples = [messages[0]["content"] + eos_token for messages in examples["prompt"]]
if not data_args.packing:
if data_args.template == "gemma":
text_examples = [tokenizer.bos_token + example for example in text_examples]
result = tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len, truncation=True)
else:
tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
block_size = data_args.cutoff_len
total_length = (total_length // block_size) * block_size
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
if data_args.template == "gemma":
for i in range(len(result["input_ids"])):
result["input_ids"][i][0] = tokenizer.bos_token_id
return result
pretrain/LLaMA-Factory/src/llamafactory/data/processors/processor_utils.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import bisect
from typing import List, Sequence, Tuple
def search_for_fit(numbers: Sequence[int], capacity: int) -> int:
r"""
Finds the index of largest number that fits into the knapsack with the given capacity.
"""
index = bisect.bisect(numbers, capacity)
return -1 if index == 0 else (index - 1)
def greedy_knapsack(numbers: List[int], capacity: int) -> List[List[int]]:
r"""
An efficient greedy algorithm with binary search for the knapsack problem.
"""
numbers.sort() # sort numbers in ascending order for binary search
knapsacks = []
while numbers:
current_knapsack = []
remaining_capacity = capacity
while True:
index = search_for_fit(numbers, remaining_capacity)
if index == -1:
break # no more numbers fit in this knapsack
remaining_capacity -= numbers[index] # update the remaining capacity
current_knapsack.append(numbers.pop(index)) # add the number to knapsack
knapsacks.append(current_knapsack)
return knapsacks
def infer_seqlen(source_len: int, target_len: int, cutoff_len: int) -> Tuple[int, int]:
r"""
Computes the real sequence length after truncation by the cutoff_len.
"""
if target_len * 2 < cutoff_len: # truncate source
max_target_len = cutoff_len
elif source_len * 2 < cutoff_len: # truncate target
max_target_len = cutoff_len - source_len
else: # truncate both
max_target_len = int(cutoff_len * (target_len / (source_len + target_len)))
new_target_len = min(max_target_len, target_len)
max_source_len = max(cutoff_len - new_target_len, 0)
new_source_len = min(max_source_len, source_len)
return new_source_len, new_target_len
pretrain/LLaMA-Factory/src/llamafactory/data/processors/supervised.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
from .processor_utils import greedy_knapsack, infer_seqlen
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..template import Template
logger = get_logger(__name__)
def _encode_supervised_example(
prompt: Sequence[Dict[str, str]],
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
train_on_prompt: bool,
mask_history: bool,
) -> Tuple[List[int], List[int]]:
messages = prompt + response
input_ids, labels = [], []
input_ids, labels = template.mm_plugin.process_token_ids(input_ids, labels, tokenizer, processor)
encoded_pairs = template.encode_multiturn(tokenizer, messages, system, tools)
total_length = 1 if template.efficient_eos else 0
if mask_history:
encoded_pairs = encoded_pairs[::-1] # high priority for last turns
for turn_idx, (source_ids, target_ids) in enumerate(encoded_pairs):
if total_length >= cutoff_len:
break
source_len, target_len = infer_seqlen(len(source_ids), len(target_ids), cutoff_len - total_length)
source_ids = source_ids[:source_len]
target_ids = target_ids[:target_len]
total_length += source_len + target_len
if train_on_prompt:
source_label = source_ids
elif template.efficient_eos:
source_label = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (source_len - 1)
else:
source_label = [IGNORE_INDEX] * source_len
if mask_history and turn_idx != 0: # train on the last turn only
target_label = [IGNORE_INDEX] * target_len
else:
target_label = target_ids
if mask_history: # reversed sequences
input_ids = source_ids + target_ids + input_ids
labels = source_label + target_label + labels
else:
input_ids += source_ids + target_ids
labels += source_label + target_label
if template.efficient_eos:
input_ids += [tokenizer.eos_token_id]
labels += [tokenizer.eos_token_id]
return input_ids, labels
def preprocess_supervised_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# build inputs with format `<bos> X Y <eos>` and labels with format `<ignore> ... <ignore> Y <eos>`
# for multiturn examples, we only mask the prompt part in each prompt-response pair.
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels = _encode_supervised_example(
prompt=prompt,
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
template=template,
tokenizer=tokenizer,
processor=processor,
cutoff_len=data_args.cutoff_len,
train_on_prompt=data_args.train_on_prompt,
mask_history=data_args.mask_history,
)
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={"token_type_ids": len(input_ids)},
processor=processor,
)
return model_inputs
def preprocess_packed_supervised_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# build inputs with format `<bos> X1 Y1 <eos> <bos> X2 Y2 <eos>`
# and labels with format `<ignore> ... <ignore> Y1 <eos> <ignore> ... <ignore> Y2 <eos>`
valid_num = 0
batch_input_ids, batch_labels = [], []
lengths = []
length2indexes = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
input_ids, labels = _encode_supervised_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
template=template,
tokenizer=tokenizer,
processor=None,
cutoff_len=data_args.cutoff_len - 1, # reserved for the padding token
train_on_prompt=data_args.train_on_prompt,
mask_history=data_args.mask_history,
)
length = len(input_ids)
if length > data_args.cutoff_len:
logger.warning("Dropped lengthy example with length {} > {}.".format(length, data_args.cutoff_len))
else:
lengths.append(length)
length2indexes[length].append(valid_num)
batch_input_ids.append(input_ids)
batch_labels.append(labels)
valid_num += 1
model_inputs = defaultdict(list)
knapsacks = greedy_knapsack(lengths, data_args.cutoff_len - 1) # reserved for the padding token
for knapsack in knapsacks:
packed_input_ids, packed_attention_masks, packed_labels = [], [], []
for i, length in enumerate(knapsack):
index = length2indexes[length].pop()
packed_input_ids += batch_input_ids[index]
packed_labels += batch_labels[index]
if data_args.neat_packing:
packed_attention_masks += [i + 1] * len(batch_input_ids[index]) # start from 1
else:
packed_attention_masks += [1] * len(batch_input_ids[index])
if len(packed_input_ids) < data_args.cutoff_len:
pad_length = data_args.cutoff_len - len(packed_input_ids)
packed_input_ids += [tokenizer.pad_token_id] * pad_length
packed_labels += [IGNORE_INDEX] * pad_length
if data_args.neat_packing:
packed_attention_masks += [0] * pad_length
else:
packed_attention_masks += [1] * pad_length # more efficient flash_attn
if len(packed_input_ids) != data_args.cutoff_len:
raise ValueError("The length of packed example should be identical to the cutoff length.")
model_inputs["input_ids"].append(packed_input_ids)
model_inputs["attention_mask"].append(packed_attention_masks)
model_inputs["labels"].append(packed_labels)
return model_inputs
def print_supervised_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
valid_labels = list(filter(lambda x: x != IGNORE_INDEX, example["labels"]))
print("input_ids:\n{}".format(example["input_ids"]))
print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
print("label_ids:\n{}".format(example["labels"]))
print("labels:\n{}".format(tokenizer.decode(valid_labels, skip_special_tokens=False)))
pretrain/LLaMA-Factory/src/llamafactory/data/processors/unsupervised.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Sequence, Tuple
from ...extras.logging import get_logger
from ..data_utils import Role
from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
from ..template import Template
logger = get_logger(__name__)
def _encode_unsupervised_example(
prompt: Sequence[Dict[str, str]],
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int]]:
if len(response) == 1:
messages = prompt + response
else:
messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
input_ids, labels = template.encode_oneturn(tokenizer, messages, system, tools)
if template.efficient_eos:
labels += [tokenizer.eos_token_id]
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, tokenizer, processor)
source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
input_ids = input_ids[:source_len]
labels = labels[:target_len]
return input_ids, labels
def preprocess_unsupervised_dataset(
examples: Dict[str, List[Any]],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
data_args: "DataArguments",
) -> Dict[str, List[Any]]:
# build inputs with format `<bos> X` and labels with format `Y <eos>`
model_inputs = defaultdict(list)
for i in range(len(examples["prompt"])):
if len(examples["prompt"][i]) % 2 != 1:
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels = _encode_unsupervised_example(
prompt=prompt,
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
template=template,
tokenizer=tokenizer,
processor=processor,
cutoff_len=data_args.cutoff_len,
)
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={"token_type_ids": len(input_ids)},
processor=processor,
)
return model_inputs
def print_unsupervised_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None:
print("input_ids:\n{}".format(example["input_ids"]))
print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
pretrain/LLaMA-Factory/src/llamafactory/data/tool_utils.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import re
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Any, Dict, List, Tuple, Union
from .data_utils import SLOTS
DEFAULT_TOOL_PROMPT = (
"You have access to the following tools:\n{tool_text}"
"Use the following format if using a tool:\n"
"```\n"
"Action: tool name (one of [{tool_names}])\n"
"Action Input: the input to the tool, in a JSON format representing the kwargs "
"""(e.g. ```{{"input": "hello world", "num_beams": 5}}```)\n"""
"```\n"
)
GLM4_TOOL_PROMPT = (
"你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，"
"你的任务是针对用户的问题和要求提供适当的答复和支持。# 可用工具{tool_text}"
)
@dataclass
class ToolUtils(ABC):
@staticmethod
@abstractmethod
def get_function_slots() -> SLOTS: ...
@staticmethod
@abstractmethod
def tool_formatter(tools: List[Dict[str, Any]]) -> str: ...
@staticmethod
@abstractmethod
def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]: ...
class DefaultToolUtils(ToolUtils):
@staticmethod
def get_function_slots() -> SLOTS:
return ["Action: {{name}}\nAction Input: {{arguments}}\n"]
@staticmethod
def tool_formatter(tools: List[Dict[str, Any]]) -> str:
tool_text = ""
tool_names = []
for tool in tools:
param_text = ""
for name, param in tool["parameters"]["properties"].items():
required, enum, items = "", "", ""
if name in tool["parameters"].get("required", []):
required = ", required"
if param.get("enum", None):
enum = ", should be one of [{}]".format(", ".join(param["enum"]))
if param.get("items", None):
items = ", where each item should be {}".format(param["items"].get("type", ""))
param_text += " - {name} ({type}{required}): {desc}{enum}{items}\n".format(
name=name,
type=param.get("type", ""),
required=required,
desc=param.get("description", ""),
enum=enum,
items=items,
)
tool_text += "> Tool Name: {name}\nTool Description: {desc}\nTool Args:\n{args}\n".format(
name=tool["name"], desc=tool.get("description", ""), args=param_text
)
tool_names.append(tool["name"])
return DEFAULT_TOOL_PROMPT.format(tool_text=tool_text, tool_names=", ".join(tool_names))
@staticmethod
def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
regex = re.compile(r"Action:\s*([a-zA-Z0-9_]+)\s*Action Input:\s*(.+?)(?=\s*Action:|\s*$)", re.DOTALL)
action_match: List[Tuple[str, str]] = re.findall(regex, content)
if not action_match:
return content
results = []
for match in action_match:
tool_name = match[0].strip()
tool_input = match[1].strip().strip('"').strip("```")
try:
arguments = json.loads(tool_input)
results.append((tool_name, json.dumps(arguments, ensure_ascii=False)))
except json.JSONDecodeError:
return content
return results
class GLM4ToolUtils(ToolUtils):
@staticmethod
def get_function_slots() -> SLOTS:
return ["{{name}}\n{{arguments}}"]
@staticmethod
def tool_formatter(tools: List[Dict[str, Any]]) -> str:
tool_text = ""
for tool in tools:
tool_text += "\n\n## {name}\n\n{body}\n在调用上述函数时，请使用 Json 格式表示调用的参数。".format(
name=tool["name"], body=json.dumps(tool, indent=4, ensure_ascii=False)
)
return GLM4_TOOL_PROMPT.format(tool_text=tool_text)
@staticmethod
def tool_extractor(content: str) -> Union[str, List[Tuple[str, str]]]:
if "\n" not in content:
return content
tool_name, tool_input = content.split("\n", maxsplit=1)
try:
arguments = json.loads(tool_input)
except json.JSONDecodeError:
return content
return [(tool_name, json.dumps(arguments, ensure_ascii=False))]
pretrain/LLaMA-Factory/tests/data/processors/test_feedback.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
import pytest
from datasets import load_dataset
from transformers import AutoTokenizer
from llamafactory.extras.constants import IGNORE_INDEX
from llamafactory.train.test_utils import load_train_dataset
DEMO_DATA = os.environ.get("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "kto",
"do_train": True,
"finetuning_type": "full",
"dataset": "kto_en_demo",
"dataset_dir": "REMOTE:" + DEMO_DATA,
"template": "llama3",
"cutoff_len": 8192,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
@pytest.mark.parametrize("num_samples", [16])
def test_feedback_data(num_samples: int):
train_dataset = load_train_dataset(**TRAIN_ARGS)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="kto_en_demo", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
messages = original_data["messages"][index]
ref_input_ids = ref_tokenizer.apply_chat_template(messages)
prompt_len = len(ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True))
ref_labels = [IGNORE_INDEX] * prompt_len + ref_input_ids[prompt_len:]
assert train_dataset["input_ids"][index] == ref_input_ids
assert train_dataset["labels"][index] == ref_labels
assert train_dataset["kto_tags"][index] == original_data["label"][index]
pretrain/LLaMA-Factory/tests/data/processors/test_pairwise.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
from typing import Dict, List
import pytest
from datasets import load_dataset
from transformers import AutoTokenizer
from llamafactory.extras.constants import IGNORE_INDEX
from llamafactory.train.test_utils import load_train_dataset
DEMO_DATA = os.environ.get("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "rm",
"do_train": True,
"finetuning_type": "full",
"dataset": "dpo_en_demo",
"dataset_dir": "REMOTE:" + DEMO_DATA,
"template": "llama3",
"cutoff_len": 8192,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
def _convert_sharegpt_to_openai(messages: List[Dict[str, str]]) -> List[Dict[str, str]]:
role_mapping = {"human": "user", "gpt": "assistant", "system": "system"}
new_messages = []
for message in messages:
new_messages.append({"role": role_mapping[message["from"]], "content": message["value"]})
return new_messages
@pytest.mark.parametrize("num_samples", [16])
def test_pairwise_data(num_samples: int):
train_dataset = load_train_dataset(**TRAIN_ARGS)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="dpo_en_demo", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
chosen_messages = original_data["conversations"][index] + [original_data["chosen"][index]]
rejected_messages = original_data["conversations"][index] + [original_data["rejected"][index]]
chosen_messages = _convert_sharegpt_to_openai(chosen_messages)
rejected_messages = _convert_sharegpt_to_openai(rejected_messages)
ref_chosen_input_ids = ref_tokenizer.apply_chat_template(chosen_messages)
chosen_prompt_len = len(ref_tokenizer.apply_chat_template(chosen_messages[:-1], add_generation_prompt=True))
ref_chosen_labels = [IGNORE_INDEX] * chosen_prompt_len + ref_chosen_input_ids[chosen_prompt_len:]
ref_rejected_input_ids = ref_tokenizer.apply_chat_template(rejected_messages)
rejected_prompt_len = len(
ref_tokenizer.apply_chat_template(rejected_messages[:-1], add_generation_prompt=True)
)
ref_rejected_labels = [IGNORE_INDEX] * rejected_prompt_len + ref_rejected_input_ids[rejected_prompt_len:]
assert train_dataset["chosen_input_ids"][index] == ref_chosen_input_ids
assert train_dataset["chosen_labels"][index] == ref_chosen_labels
assert train_dataset["rejected_input_ids"][index] == ref_rejected_input_ids
assert train_dataset["rejected_labels"][index] == ref_rejected_labels
pretrain/LLaMA-Factory/tests/data/processors/test_processor_utils.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Tuple
import pytest
from llamafactory.data.processors.processor_utils import infer_seqlen
@pytest.mark.parametrize(
"test_input,test_output",
[
((3000, 2000, 1000), (600, 400)),
((2000, 3000, 1000), (400, 600)),
((1000, 100, 1000), (900, 100)),
((100, 1000, 1000), (100, 900)),
((100, 500, 1000), (100, 500)),
((500, 100, 1000), (500, 100)),
((10, 10, 1000), (10, 10)),
],
)
def test_infer_seqlen(test_input: Tuple[int, int, int], test_output: Tuple[int, int]):
assert test_output == infer_seqlen(*test_input)
pretrain/LLaMA-Factory/tests/data/processors/test_supervised.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
import pytest
from datasets import load_dataset
from transformers import AutoTokenizer
from llamafactory.extras.constants import IGNORE_INDEX
from llamafactory.train.test_utils import load_train_dataset
DEMO_DATA = os.environ.get("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_DATA = os.environ.get("TINY_DATA", "llamafactory/tiny-supervised-dataset")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "sft",
"do_train": True,
"finetuning_type": "full",
"template": "llama3",
"cutoff_len": 8192,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
@pytest.mark.parametrize("num_samples", [16])
def test_supervised_single_turn(num_samples: int):
train_dataset = load_train_dataset(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(TINY_DATA, split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
prompt = original_data["instruction"][index]
if original_data["input"][index]:
prompt += "\n" + original_data["input"][index]
messages = [
{"role": "user", "content": prompt},
{"role": "assistant", "content": original_data["output"][index]},
]
ref_input_ids = ref_tokenizer.apply_chat_template(messages)
assert train_dataset["input_ids"][index] == ref_input_ids
@pytest.mark.parametrize("num_samples", [8])
def test_supervised_multi_turn(num_samples: int):
train_dataset = load_train_dataset(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
ref_input_ids = ref_tokenizer.apply_chat_template(original_data["messages"][index])
assert train_dataset["input_ids"][index] == ref_input_ids
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_train_on_prompt(num_samples: int):
train_dataset = load_train_dataset(
dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", train_on_prompt=True, **TRAIN_ARGS
)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
ref_ids = ref_tokenizer.apply_chat_template(original_data["messages"][index])
assert train_dataset["input_ids"][index] == ref_ids
assert train_dataset["labels"][index] == ref_ids
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_mask_history(num_samples: int):
train_dataset = load_train_dataset(
dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", mask_history=True, **TRAIN_ARGS
)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
messages = original_data["messages"][index]
ref_input_ids = ref_tokenizer.apply_chat_template(messages)
prompt_len = len(ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True))
ref_label_ids = [IGNORE_INDEX] * prompt_len + ref_input_ids[prompt_len:]
assert train_dataset["input_ids"][index] == ref_input_ids
assert train_dataset["labels"][index] == ref_label_ids
pretrain/LLaMA-Factory/tests/data/processors/test_unsupervised.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
import pytest
from datasets import load_dataset
from transformers import AutoTokenizer
from llamafactory.train.test_utils import load_train_dataset
DEMO_DATA = os.environ.get("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_DATA = os.environ.get("TINY_DATA", "llamafactory/tiny-supervised-dataset")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "ppo",
"do_train": True,
"finetuning_type": "full",
"reward_model": "",
"reward_model_type": "full",
"dataset": "system_chat",
"dataset_dir": "REMOTE:" + DEMO_DATA,
"template": "llama3",
"cutoff_len": 8192,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
@pytest.mark.parametrize("num_samples", [16])
def test_unsupervised_data(num_samples: int):
train_dataset = load_train_dataset(**TRAIN_ARGS)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
messages = original_data["messages"][index]
ref_ids = ref_tokenizer.apply_chat_template(messages)
ref_input_ids = ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True)
ref_labels = ref_ids[len(ref_input_ids) :]
assert train_dataset["input_ids"][index] == ref_input_ids
assert train_dataset["labels"][index] == ref_labels
pretrain/LLaMA-Factory/tests/data/test_collator.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from llamafactory.data.collator import prepare_4d_attention_mask
def test_4d_attention_mask():
o = 0.0
x = torch.finfo(torch.float16).min
attention_mask_with_indices = torch.tensor(
[
[1, 1, 2, 2, 2, 0],
[1, 2, 2, 3, 3, 3],
]
)
attention_mask_computed = prepare_4d_attention_mask(attention_mask_with_indices, torch.float16)
attention_mask_expected = torch.tensor(
[
[
[
[o, x, x, x, x, x],
[o, o, x, x, x, x],
[x, x, o, x, x, x],
[x, x, o, o, x, x],
[x, x, o, o, o, x],
[x, x, x, x, x, x],
]
],
[
[
[o, x, x, x, x, x],
[x, o, x, x, x, x],
[x, o, o, x, x, x],
[x, x, x, o, x, x],
[x, x, x, o, o, x],
[x, x, x, o, o, o],
]
],
],
dtype=torch.float16,
)
assert list(attention_mask_computed.size()) == [2, 1, 6, 6]
assert torch.all(attention_mask_computed == attention_mask_expected)
pretrain/LLaMA-Factory/tests/data/test_formatter.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
from llamafactory.data.formatter import EmptyFormatter, FunctionFormatter, StringFormatter, ToolFormatter
def test_empty_formatter():
formatter = EmptyFormatter(slots=["\n"])
assert formatter.apply() == ["\n"]
def test_string_formatter():
formatter = StringFormatter(slots=["<s>", "Human: {{content}}\nAssistant:"])
assert formatter.apply(content="Hi") == ["<s>", "Human: Hi\nAssistant:"]
def test_function_formatter():
formatter = FunctionFormatter(slots=[], tool_format="default")
tool_calls = json.dumps({"name": "tool_name", "arguments": {"foo": "bar", "size": 10}})
assert formatter.apply(content=tool_calls) == [
"""Action: tool_name\nAction Input: {\"foo\": \"bar\", \"size\": 10}\n"""
]
def test_multi_function_formatter():
formatter = FunctionFormatter(slots=[], tool_format="default")
tool_calls = json.dumps([{"name": "tool_name", "arguments": {"foo": "bar", "size": 10}}] * 2)
assert formatter.apply(content=tool_calls) == [
"""Action: tool_name\nAction Input: {\"foo\": \"bar\", \"size\": 10}\n""",
"""Action: tool_name\nAction Input: {\"foo\": \"bar\", \"size\": 10}\n""",
]
def test_default_tool_formatter():
formatter = ToolFormatter(tool_format="default")
tools = [
{
"name": "test_tool",
"description": "tool_desc",
"parameters": {
"type": "object",
"properties": {
"foo": {"type": "string", "description": "foo_desc"},
"bar": {"type": "number", "description": "bar_desc"},
},
"required": ["foo"],
},
}
]
assert formatter.apply(content=json.dumps(tools)) == [
"You have access to the following tools:\n"
"> Tool Name: test_tool\n"
"Tool Description: tool_desc\n"
"Tool Args:\n"
" - foo (string, required): foo_desc\n"
" - bar (number): bar_desc\n\n"
"Use the following format if using a tool:\n"
"```\n"
"Action: tool name (one of [test_tool])\n"
"Action Input: the input to the tool, in a JSON format representing the kwargs "
"""(e.g. ```{"input": "hello world", "num_beams": 5}```)\n"""
"```\n"
]
def test_default_tool_extractor():
formatter = ToolFormatter(tool_format="default")
result = """Action: test_tool\nAction Input: {"foo": "bar", "size": 10}\n"""
assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
def test_default_multi_tool_extractor():
formatter = ToolFormatter(tool_format="default")
result = (
"""Action: test_tool\nAction Input: {"foo": "bar", "size": 10}\n"""
"""Action: another_tool\nAction Input: {"foo": "job", "size": 2}\n"""
)
assert formatter.extract(result) == [
("test_tool", """{"foo": "bar", "size": 10}"""),
("another_tool", """{"foo": "job", "size": 2}"""),
]
def test_glm4_tool_formatter():
formatter = ToolFormatter(tool_format="glm4")
tools = [
{
"name": "test_tool",
"description": "tool_desc",
"parameters": {
"type": "object",
"properties": {
"foo": {"type": "string", "description": "foo_desc"},
"bar": {"type": "number", "description": "bar_desc"},
},
"required": ["foo"],
},
}
]
assert formatter.apply(content=json.dumps(tools)) == [
"你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，"
"你的任务是针对用户的问题和要求提供适当的答复和支持。# 可用工具\n\n"
"## test_tool\n\n{}\n在调用上述函数时，请使用 Json 格式表示调用的参数。".format(json.dumps(tools[0], indent=4))
]
def test_glm4_tool_extractor():
formatter = ToolFormatter(tool_format="glm4")
result = """test_tool\n{"foo": "bar", "size": 10}\n"""
assert formatter.extract(result) == [("test_tool", """{"foo": "bar", "size": 10}""")]
pretrain/LLaMA-Factory/tests/data/test_template.py 0 → 100644
# Copyright 2024 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from typing import TYPE_CHECKING, List, Sequence
import pytest
from transformers import AutoTokenizer
from llamafactory.data import get_template_and_fix_tokenizer
if TYPE_CHECKING:
from transformers import PreTrainedTokenizer
HF_TOKEN = os.environ.get("HF_TOKEN", None)
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
MESSAGES = [
{"role": "user", "content": "How are you"},
{"role": "assistant", "content": "I am fine!"},
{"role": "user", "content": "你好"},
{"role": "assistant", "content": "很高兴认识你！"},
]
def _check_tokenization(
tokenizer: "PreTrainedTokenizer", batch_input_ids: Sequence[Sequence[int]], batch_text: Sequence[str]
) -> None:
for input_ids, text in zip(batch_input_ids, batch_text):
assert input_ids == tokenizer.encode(text, add_special_tokens=False)
assert tokenizer.decode(input_ids) == text
def _check_single_template(
model_id: str, template_name: str, prompt_str: str, answer_str: str, extra_str: str, use_fast: bool
) -> List[str]:
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=use_fast, token=HF_TOKEN)
content_str = tokenizer.apply_chat_template(MESSAGES, tokenize=False)
content_ids = tokenizer.apply_chat_template(MESSAGES, tokenize=True)
template = get_template_and_fix_tokenizer(tokenizer, name=template_name)
prompt_ids, answer_ids = template.encode_oneturn(tokenizer, MESSAGES)
assert content_str == prompt_str + answer_str + extra_str
assert content_ids == prompt_ids + answer_ids + tokenizer.encode(extra_str, add_special_tokens=False)
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
return content_ids
def _check_template(model_id: str, template_name: str, prompt_str: str, answer_str: str, extra_str: str = "") -> None:
"""
Checks template for both the slow tokenizer and the fast tokenizer.
Args:
model_id: the model id on hugging face hub.
template_name: the template name.
prompt_str: the string corresponding to the prompt part.
answer_str: the string corresponding to the answer part.
extra_str: the extra string in the jinja template of the original tokenizer.
"""
slow_ids = _check_single_template(model_id, template_name, prompt_str, answer_str, extra_str, use_fast=False)
fast_ids = _check_single_template(model_id, template_name, prompt_str, answer_str, extra_str, use_fast=True)
assert slow_ids == fast_ids
@pytest.mark.parametrize("use_fast", [True, False])
def test_encode_oneturn(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
template = get_template_and_fix_tokenizer(tokenizer, name="llama3")
prompt_ids, answer_ids = template.encode_oneturn(tokenizer, MESSAGES)
prompt_str = (
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\nI am fine!<|eot_id|>"
"<|start_header_id|>user<|end_header_id|>\n\n你好<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
answer_str = "很高兴认识你！<|eot_id|>"
_check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
@pytest.mark.parametrize("use_fast", [True, False])
def test_encode_multiturn(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
template = get_template_and_fix_tokenizer(tokenizer, name="llama3")
encoded_pairs = template.encode_multiturn(tokenizer, MESSAGES)
prompt_str_1 = (
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
answer_str_1 = "I am fine!<|eot_id|>"
prompt_str_2 = (
"<|start_header_id|>user<|end_header_id|>\n\n你好<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
answer_str_2 = "很高兴认识你！<|eot_id|>"
_check_tokenization(
tokenizer,
(encoded_pairs[0][0], encoded_pairs[0][1], encoded_pairs[1][0], encoded_pairs[1][1]),
(prompt_str_1, answer_str_1, prompt_str_2, answer_str_2),
)
@pytest.mark.parametrize("use_fast", [True, False])
def test_jinja_template(use_fast: bool):
tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
get_template_and_fix_tokenizer(tokenizer, name="llama3")
assert tokenizer.chat_template != ref_tokenizer.chat_template
assert tokenizer.apply_chat_template(MESSAGES) == ref_tokenizer.apply_chat_template(MESSAGES)
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_gemma_template():
prompt_str = (
"<bos><start_of_turn>user\nHow are you<end_of_turn>\n"
"<start_of_turn>model\nI am fine!<end_of_turn>\n"
"<start_of_turn>user\n你好<end_of_turn>\n"
"<start_of_turn>model\n"
)
answer_str = "很高兴认识你！"
_check_template("google/gemma-2-9b-it", "gemma", prompt_str, answer_str, extra_str="<end_of_turn>\n")
@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
def test_llama3_template():
prompt_str = (
"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\nI am fine!<|eot_id|>"
"<|start_header_id|>user<|end_header_id|>\n\n你好<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
answer_str = "很高兴认识你！<|eot_id|>"
_check_template("meta-llama/Meta-Llama-3-8B-Instruct", "llama3", prompt_str, answer_str)
def test_qwen_template():
prompt_str = (
"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
"<|im_start|>user\nHow are you<|im_end|>\n"
"<|im_start|>assistant\nI am fine!<|im_end|>\n"
"<|im_start|>user\n你好<|im_end|>\n"
"<|im_start|>assistant\n"
)
answer_str = "很高兴认识你！<|im_end|>"
_check_template("Qwen/Qwen2-7B-Instruct", "qwen", prompt_str, answer_str, extra_str="\n")
@pytest.mark.skip(reason="The fast tokenizer of Yi model is corrupted.")
def test_yi_template():
prompt_str = (
"<|im_start|>user\nHow are you<|im_end|>\n"
"<|im_start|>assistant\nI am fine!<|im_end|>\n"
"<|im_start|>user\n你好<|im_end|>\n"
"<|im_start|>assistant\n"
)
answer_str = "很高兴认识你！<|im_end|>"
_check_template("01-ai/Yi-1.5-6B-Chat", "yi", prompt_str, answer_str)
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment