fix some bugs

README.md

@@ -222,6 +222,12 @@ builtins.print = traced_print
 'event': 'stdin pipe closed early while writing large input', 'logger': 'sandbox.runners.base', 'level': 'warning'
 ```
 
+#### Prompt过滤长度没加tool导致后续prompt长度超了报错
+
+verl在`verl/utils/dataset/rl_dataset.py`的`_read_files_and_tokenize`函数里面初筛prompt长度的时候没有把tool对应的prompt给算上，导致prompt长度估计短了，在后续运行的时候会因为长度超了报错。
+
+
+
 #### 最后贴一个（应该是）正常的初始测试集准确率
 
 **警告：这个结果可能被sympy超时影响导致acc偏低！！！**

SandboxFusion/sandbox/runners/base.py

@@ -72,7 +72,8 @@ async def run_command_bare(command: str | List[str],
                 if p.stdin:
                     print('p.stdin is', p.stdin)
                     p.stdin.write(stdin.encode())
-                    p.stdin.flush()
+                    # p.stdin.flush()
+                    await p.stdin.drain()
                 else:
                     logger.warning("Attempted to write to stdin, but stdin is closed.")
             except Exception as e:

examples/data_preprocess/codev.py

@@ -96,14 +96,6 @@ if __name__ == '__main__':
         question = mk_prompt_r1(question)
         return question
 
-    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path)
-    raw_dataset = raw_dataset.shuffle(seed=42)
-    print(len(raw_dataset))
-
-    assert len(raw_dataset) >= TRAIN_SIZE + TEST_SIZE
-    train_dataset = raw_dataset.select(range(TRAIN_SIZE))
-    test_dataset = raw_dataset.select(range(TRAIN_SIZE, TRAIN_SIZE + TEST_SIZE))
-
     def make_map_fn(split):
         def process_fn(example, idx):
             if "ground_truth" in example:
@@ -134,8 +126,38 @@ if __name__ == '__main__':
             return data
         return process_fn
 
-    train_dataset = train_dataset.map(function=make_map_fn('train'), with_indices=True)
-    test_dataset = test_dataset.map(function=make_map_fn('test'), with_indices=True)
+    raw_dataset = raw_dataset.map(function=make_map_fn('train'), with_indices=True)
+
+    MAX_TOKEN_LEN = 2048
+    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path)
+    from omegaconf import OmegaConf
+    tools_config_file = 'examples/tir/sandbox_fusion_verilog_config.yaml'
+    tools = OmegaConf.load(tools_config_file)
+    tools = [OmegaConf.to_container(tools.tools[0].tool_schema, resolve=True)]
+    
+    def filter_by_token_len(example):
+        question = make_question(example["prompt"])
+        token_ids = tokenizer.apply_chat_template(
+            question,
+            tools=tools,
+            add_generation_prompt=False,
+            tokenize=True,
+            truncation=False,
+        )
+        token_len = len(token_ids)
+        example["question_token_len"] = token_len
+        return token_len <= MAX_TOKEN_LEN - 20
+
+    raw_dataset = raw_dataset.filter(
+        function=filter_by_token_len,
+        desc="Filtering train dataset by token length"  # 显示进度描述
+    )
+    raw_dataset = raw_dataset.shuffle(seed=42)
+    print(len(raw_dataset))
+
+    assert len(raw_dataset) >= TRAIN_SIZE + TEST_SIZE
+    train_dataset = raw_dataset.select(range(TRAIN_SIZE))
+    test_dataset = raw_dataset.select(range(TRAIN_SIZE, TRAIN_SIZE + TEST_SIZE))
 
     local_dir = args.local_dir
     hdfs_dir = args.hdfs_dir

examples/tir/tencent/run_tir_codev.sh

@@ -3,14 +3,15 @@ export VERL_LOGGING_LEVEL=INFO
 # export VERL_LOGGING_LEVEL=DEBUG
 # MODEL_PATH=/nfs_global/models/Qwen2.5-Coder-7B-Instruct
 MODEL_PATH=/nfs_global/models/Qwen3-8B
+DATA_PATH=$CURR_DIR/data/codev/v1/16k_r1_filtered
 
 python3 -X faulthandler -u -m verl.trainer.main_ppo \
     reward_model.sandbox_fusion.url=$SANDBOX_URL \
     reward_model.sandbox_fusion.max_concurrent=128 \
     reward_model.reward_manager=prime \
     algorithm.adv_estimator=grpo \
-    data.train_files=$CURR_DIR/data/codev/v1/16k_r1_filtered/train.parquet \
-    data.val_files=$CURR_DIR/data/codev/v1/16k_r1_filtered/test.parquet \
+    data.train_files=$DATA_PATH/train.parquet \
+    data.val_files=$DATA_PATH/test.parquet \
     data.train_batch_size=64 \
     data.max_prompt_length=2048 \
     data.max_response_length=8192 \
@@ -44,6 +45,7 @@ python3 -X faulthandler -u -m verl.trainer.main_ppo \
     trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=20 \
+    trainer.default_local_dir=$SAVE_DIR \
     trainer.test_freq=1 \
     trainer.total_epochs=15 $@
 

examples/tir/tencent/run_tir_example.sh

@@ -41,6 +41,7 @@ python3 -X faulthandler -u -m verl.trainer.main_ppo \
     trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=20 \
+    trainer.default_local_dir=$SAVE_DIR \
     trainer.test_freq=1 \
     trainer.total_epochs=15 $@
 

scripts/preprocess.sh

@@ -13,4 +13,5 @@
 # python examples/data_preprocess/codev.py --data_path /nfs_global/S/lvhanqi/codev_data/sft_model_87k_correct1234_filter_qwen7b32b_data.jsonl --local_dir data/codev/v1/qwen7b32b_filter_gt_r1_14k --gt r1 --train_size 14654 --test_size 300
 
 # tencent cloud
-python examples/data_preprocess/codev.py --tool --tokenizer_path /nfs_global/models/Qwen2.5-Coder-7B-Instruct/ --data_path /nfs_global/datasets/codev/codev_r1_rl_16k.jsonl --local_dir data/codev/v1/16k_r1_filtered --train_size 15691 --test_size 300
\ No newline at end of file
+python examples/data_preprocess/codev.py --tool --tokenizer_path /nfs_global/models/Qwen2.5-Coder-7B-Instruct/ --data_path /nfs_global/datasets/codev/codev_r1_rl_16k.jsonl --local_dir data/codev/v1/16k_r1_filtered --train_size 15685 --test_size 300
+# python examples/data_preprocess/codev.py --tool --tokenizer_path /nfs_global/models/Qwen2.5-Coder-7B-Instruct/ --data_path /nfs_global/datasets/codev/codev_r1_rl_16k.jsonl --local_dir data/codev/v1/test --train_size 139 --test_size 11
\ No newline at end of file

verl/tools/sandbox_fusion_tools.py

@@ -159,7 +159,7 @@ class SandboxFusionTool(BaseTool):
             code = str(code)
 
         result = await self.execution_pool.execute.remote(self.execute_code, instance_id, code, timeout, language)
-        print(f'In sandbox code execution, code is {code}, timeout is {timeout}, language is {language}, result is {result}, parameter keys are {parameters.keys()}.')
+        # print(f'In sandbox code execution, code is {code}, timeout is {timeout}, language is {language}, result is {result}, parameter keys are {parameters.keys()}.')
 
         return result, result, result.strip()