fix bugs

codecritic/cli/eval.py

 import argparse
+from collections import defaultdict
+import json
 from functools import partial
-from pathlib import Path
 
 import codecritic.evaluation.metric as metric
-from codecritic.utils.json import load_jsonl, save_jsonl
-
-def eval(samples_path):
-    model, testset = samples_path.stem.split('-')[:2]
-    def f(item):
-        item["model"] = model
-        item["testset"] = testset
-    
-    samples = load_jsonl(samples_path)
+from codecritic.utils.json import load_jsonl
 
+def eval(scores):
     ks = list(range(1, 17))
 
     results = []
-    results.append(metric.pass_at_k(samples, ks))
-    results.append(metric.top_at_k(samples, ks, metric.postive_and_negative))
-    results.append(metric.top_at_k(samples, ks, metric.positive_only))
+    results.extend(metric.pass_at_k(scores, ks))
+    results.extend(metric.pass_at_k(scores, [50]))
+    results.extend(metric.top_at_k(scores, ks, metric.positive_only))
+
+    if "negative_score" in scores[0]:
+        results.extend(metric.top_at_k(scores, ks, metric.postive_and_negative))
 
-    for i in range(4):
-        threshold = 0.5 + i * 0.1
-        score_func = partial(metric.pos_neg_filter_uncertain, threshold=threshold)
-        results.append(metric.top_at_k(samples, ks, score_func))
+        for i in range(4):
+            threshold = 0.5 + i * 0.1
+            score_func = partial(metric.pos_neg_filter_uncertain, threshold=threshold)
+            results.extend(metric.top_at_k(scores, ks, score_func))
 
-    return list(map(f, results))
+    return results
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--sample_dir",
-        type=str,
-        default=None,
-        help="Path to the directory containing samples. If not provided, cached results will be used."
-    )
-    parser.add_argument("--out_dir", type=str, help="path/to/output_dir")
-    parser.add_argument(
-        "--score_func",
-        type=str,
-        default="all",
-        choices=["all", "posonly", "posneg", "posneg_filter"],  # Add valid options
-        help="Select the scoring function to use. Default: 'all'."
-    )
-    parser.add_argument("--plot", type=str, help="path/to/plot")
+    parser.add_argument("--score", type=str, help="path/to/score")
     args = parser.parse_args()
 
-    outdir = Path(args.out_dir)
-    if args.sample_dir:
-        for samples_path in Path(args.sample_dir).glob("*.jsonl"):
-            out_path = outdir / (samples_path.stem + "-eval.jsonl")
-            if not out_path.exists():
-                eval_results = eval(samples_path)
-                save_jsonl(eval_results, out_path)
-    
-    for out_path in outdir.glob("*.jsonl"):
-        pass
+    scores = load_jsonl(args.score)
+    groups = defaultdict(list)
+    for item in scores:
+        groups[item["dataset"]].append(item)
+
+    for dataset, lst in groups.items():
+        results = eval(lst)
+        for r in results:
+            r["dataset"] = dataset
+            print(json.dumps(r))
\ No newline at end of file

codecritic/cli/test_genrm.py

 import argparse
+from collections import defaultdict
 from functools import partial
 import os
 
 from transformers import AutoTokenizer
 from vllm import SamplingParams
 
-from codecritic.dataset.genrm_prompt import THINK_MESSAGE, JUDGE_MESSAGE, JUDGE_TOEKNS
+from codecritic.dataset.genrm_prompt import JUDGE_MESSAGE, JUDGE_TOEKNS
 from codecritic.utils.inference import generate_worker, score_worker
 from codecritic.utils.parallel import model_map
 from codecritic.utils.json import load_jsonl, save_jsonl
+import codecritic.dataset.algolr_prompt as algolr_prompt
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--model", type=str, help="path/to/model")
-    parser.add_argument("--sample", type=str, help="path/to/sample")
+    parser.add_argument("--testset", type=str, help="path/to/testset")
     parser.add_argument("--output", type=str, help="path/to/score")
     parser.add_argument("--reasoning", action="store_true", help="enable reasoning")
     parser.add_argument(
@@ -31,11 +33,24 @@ if __name__ == "__main__":
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
     tokenizer = AutoTokenizer.from_pretrained(args.model)
 
-    dataset = load_jsonl(args.sample)
+    dataset = load_jsonl(args.testset)
+    ds = defaultdict(list)
+    for item in dataset:
+        ds[item["task_id"]].append(item)
+
+    unsolvable = []
+    dataset = []
+    for task_id, items in ds.items():
+        if all([not x["pass"] for x in items]):
+            for item in items:
+                item["positive_score"] = 0
+                item["negative_score"] = 0
+            unsolvable.extend(items)
+        else:
+            dataset.extend(items)
 
     if args.reasoning:
-        for item in dataset:
-            item["messages"].append(THINK_MESSAGE)
+        dataset = [algolr_prompt.process_to_reason_prompt(x, None) for x in dataset]
 
         sampling_params = SamplingParams(
             n=1,
@@ -68,4 +83,4 @@ if __name__ == "__main__":
     )
     dataset = model_map(worker, dataset, args.tp)
 
-    save_jsonl(dataset, args.output)
+    save_jsonl(dataset + unsolvable, args.output)

codecritic/cli/test_orm.py

 import argparse
+from collections import defaultdict
 import json
 import requests
 from tqdm import tqdm
@@ -21,18 +22,28 @@ def get_rewards_from_server(server_url: str, messages: list[str]):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--model", type=str, help="path/to/model")
-    parser.add_argument("--sample", type=str, help="path/to/sample")
+    parser.add_argument("--testset", type=str, help="path/to/testset")
     parser.add_argument("--output", type=str, help="path/to/score")
     args = parser.parse_args()
 
     # compute score
-    dataset = load_jsonl(args.sample)
+    dataset = load_jsonl(args.testset)
+    ds = defaultdict(list)
+    for item in dataset:
+        ds[item["task_id"]].append(item)
+
+    for task_id, items in ds.items():
+        if all([not x["pass"] for x in items]):
+            for item in items:
+                item["positive_score"] = 0
+
     server_url = "http://0.0.0.0:5000/get_reward"
     tokenizer = AutoTokenizer.from_pretrained(args.model)
-    
+
     for item in tqdm(dataset):
-        query = tokenizer.apply_chat_template(item["messages"], tokenize=False)
-        score = get_rewards_from_server(server_url, [query])[0]
-        item["score"] = score
+        if 'positive_score' not in item:
+            query = tokenizer.apply_chat_template(item["messages"], tokenize=False)
+            score = get_rewards_from_server(server_url, [query])[0]
+            item["positive_score"] = score
 
     save_jsonl(dataset, args.output)

scripts/algolr.sh

@@ -13,18 +13,20 @@ apps="/lustre/S/nanziyuan/datasets/apps/"
 sft="${project}/data/train/${modelname}-sft.jsonl"
 
 ftmodel="${project}/model/qwen25_coder_inst_7b-algolr"
+testset="/lustre/S/nanziyuan/projects/ccc/data/test/qwen25_coder_inst-apps-test.jsonl"
+evalresults="/lustre/S/nanziyuan/projects/ccc/data/eval/qwen25_code_inst-apps-test-algolr-score.jsonl"
 
-### export CUDA_VISIBLE_DEVICES=0,1,2,3
+# export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 
-python -m codecritic.cli.algolr \
-       --model ${model} \
-       --dataset ${trainset} \
-       --pairinfo ${train_selected_pairs} \
-       --apps ${apps} \
-       --output ${sft} \
-       --level beginner \
-       --tp 1
+# python -m codecritic.cli.algolr \
+#        --model ${model} \
+#        --dataset ${trainset} \
+#        --pairinfo ${train_selected_pairs} \
+#        --apps ${apps} \
+#        --output ${sft} \
+#        --level beginner \
+#        --tp 1
 
 
 deepspeed --module \
@@ -50,3 +52,11 @@ openrlhf.cli.train_sft \
    --load_checkpoint \
    --gradient_checkpointing \
    --use_tensorboard "${ftmodel}_log"
+
+
+python -m codecritic.cli.test_genrm \
+       --model ${model} \
+       --testset ${testset} \
+       --output ${evalresults} \
+       --reasoning \
+       --tp 1

scripts/train_orm.sh

@@ -6,66 +6,69 @@ project="/lustre/S/nanziyuan/projects/ccc"
 model="/lustre/S/huangdi/open_for_out/models/Qwen2.5-Coder-7B-Instruct/"
 ftmodel="${project}/model/qwen25_coder_inst_7b-orm"
 
-deepspeed --module \
-openrlhf.cli.train_rm \
-   --save_path ${ftmodel} \
-   --save_steps -1 \
-   --logging_steps 1 \
-   --eval_steps -1 \
-   --train_batch_size 256 \
-   --micro_train_batch_size 1 \
-   --pretrain ${model} \
-   --bf16 \
-   --max_epochs 1 \
-   --max_len 8192 \
-   --zero_stage 3 \
-   --learning_rate 9e-6 \
-   --dataset  ${dataset} \
-   --apply_chat_template \
-   --prompt_key messages \
-   --chosen_key chosen \
-   --rejected_key rejected \
-   --flash_attn \
-   --load_checkpoint \
-   --gradient_checkpointing \
-   --use_tensorboard "${ftmodel}_log"
+testset="/lustre/S/nanziyuan/projects/ccc/data/test/qwen25_coder_inst-apps-test.jsonl"
+evalresults="/lustre/S/nanziyuan/projects/ccc/data/eval/qwen25_code_inst-apps-test-orm-score.jsonl"
 
+# deepspeed --module \
+# openrlhf.cli.train_rm \
+#    --save_path ${ftmodel} \
+#    --save_steps -1 \
+#    --logging_steps 1 \
+#    --eval_steps -1 \
+#    --train_batch_size 256 \
+#    --micro_train_batch_size 1 \
+#    --pretrain ${model} \
+#    --bf16 \
+#    --max_epochs 1 \
+#    --max_len 8192 \
+#    --zero_stage 3 \
+#    --learning_rate 9e-6 \
+#    --dataset  ${dataset} \
+#    --apply_chat_template \
+#    --prompt_key messages \
+#    --chosen_key chosen \
+#    --rejected_key rejected \
+#    --flash_attn \
+#    --load_checkpoint \
+#    --gradient_checkpointing \
+#    --use_tensorboard "${ftmodel}_log"
 
-# start_server() {
-#     echo "Starting server..."
-#     CUDA_VISIBLE_DEVICES=0 \
-#         python -m openrlhf.cli.serve_rm \
-#         --reward_pretrain ${model} \
-#         --normalize_reward \
-#         --port 5000 \
-#         --bf16 \
-#         --max_len 8192 &
-#     SERVER_PID=$!
-#     echo "Server started with PID: $SERVER_PID"
-# }
 
-# # Function to start the client
-# start_client() {
-#     echo "Starting client..."
-#     python -m codecritic.cli.run_rm_test \
-#            --model ${model} \
-#            --test "${datasets}/sample/min_test.jsonl" \
-#            --apps /lustre/S/nanziyuan/datasets/apps/
-#     CLIENT_EXIT_CODE=$?
-#     echo "Client finished with exit code: $CLIENT_EXIT_CODE"
-# }
+start_server() {
+    echo "Starting server..."
+    CUDA_VISIBLE_DEVICES=0 \
+        python -m openrlhf.cli.serve_rm \
+        --reward_pretrain ${ftmodel} \
+        --normalize_reward \
+        --port 5000 \
+        --bf16 \
+        --max_len 8192 &
+    SERVER_PID=$!
+    echo "Server started with PID: $SERVER_PID"
+}
 
-# # Function to stop the server
-# stop_server() {
-#     echo "Stopping server..."
-#     kill -SIGINT $SERVER_PID
-#     wait $SERVER_PID 2>/dev/null
-#     echo "Server stopped."
-# }
+# Function to start the client
+start_client() {
+    echo "Starting client..."
+    python -m codecritic.cli.test_orm \
+           --model ${ftmodel} \
+           --testset ${testset} \
+           --output ${evalresults}
+    CLIENT_EXIT_CODE=$?
+    echo "Client finished with exit code: $CLIENT_EXIT_CODE"
+}
 
-# start_server
-# # Give the server some time to initialize (optional)
-# sleep 60
-# start_client
-# stop_server
-# echo "Execution complete."
+# Function to stop the server
+stop_server() {
+    echo "Stopping server..."
+    kill -SIGINT $SERVER_PID
+    wait $SERVER_PID 2>/dev/null
+    echo "Server stopped."
+}
+
+start_server
+# Give the server some time to initialize (optional)
+sleep 60
+start_client
+stop_server
+echo "Execution complete."