reformat dataset

README.md

+# CodeCritic
+
+## Evaluation
+
+### APPS-500 (Top@k)
+
+### HumanEval (Top@k)
+
+### MBPP (Top@k)
+
+### HumanEvalPack-Python (Pairwise)
+
+### BigCodeBench (TODO)
+
+### Dataset Format
+
+```json
+{
+    "dataset": "string, name or identifier of the dataset",
+    "task_id": "integer or string, identifier for the task",
+    "solution_id": "integer or string, unique identifier for the solution",
+    "pass": "boolean, indicates whether the solution passed the task",
+    "skip": "boolean, set to True if no solution passes this task",
+    "messages": "list of dictionaries, conversation messages in OpenAI format",
+    "positive_score": "float, probability of the 'Yes' token",
+    "negative_score": "float, probability of the 'No' token",
+    "meta_***": "any additional data or custom fields",
+}
+```
+
+### 
\ No newline at end of file

codecritic/cli/convert_pairwise_to_sft.py

-import argparse
-from codecritic.dataset.utils import mk_message, save_jsonl_dataset
-from codecritic.utils.json import load_jsonl
-from codecritic.dataset.verify import mk_critic_verify
-
-
-def mk_sft_dataset(messages):
-    question = messages[:-1]
-    response = messages[-1:]
-    return dict(question=question, response=response)
-
-
-def convert_preference_to_sft(item):
-    message = item["messages"]
-    chosen = item["chosen"]
-    rejected = item["rejected"]
-
-    messages1 = message + chosen + mk_critic_verify(True)
-    messages2 = message + rejected + mk_critic_verify(False)
-    return mk_sft_dataset(messages1), mk_sft_dataset(messages2)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--preference_dataset", type=str)
-    parser.add_argument("--output_dir", type=str)
-    parser.add_argument("--split", type=str, default="train")
-    args = parser.parse_args()
-
-    preference_dataset = load_jsonl(args.preference_dataset)
-
-    sft_dataset = []
-    for item in preference_dataset:
-        sft_dataset.extend(convert_preference_to_sft(item))
-
-    save_jsonl_dataset(sft_dataset, args.output_dir, args.split)

codecritic/cli/mk_cov_from_api.py → codecritic/cli/critic.py

 import argparse
+from itertools import chain
+
+from codecritic.dataset.cov import (
+    convert_preference_to_vot_prompt,
+    convert_cov_to_cov_dataset,
+)
+
+from codecritic.utils.json import load_json
+from codecritic.dataset.utils import save_jsonl_dataset
+from codecritic.utils.vllm import vllm_chatcomplete
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str)
+    parser.add_argument("--preference_dataset", type=str)
+    parser.add_argument("--output_dir", type=str)
+    args = parser.parse_args()
+
+    preference_dataset = load_json(args.preference_dataset)
+    cov_prompts = [convert_preference_to_vot_prompt(x) for x in preference_dataset]
+    cov_prompts = list(chain(*cov_prompts))
+
+    sampling_params = dict(n=1, temperature=0.0, max_tokens=2048)
+    covs = vllm_chatcomplete(args.model, cov_prompts, sampling_params)
+    dataset = list(map(convert_cov_to_cov_dataset, covs))
+
+    save_jsonl_dataset(dataset, args.output_dir)
+
+import argparse
 import datetime
 from itertools import chain
 from functools import partial

codecritic/cli/mk_cov_dataset.py

-import argparse
-from itertools import chain
-
-from codecritic.dataset.cov import (
-    convert_preference_to_vot_prompt,
-    convert_cov_to_cov_dataset,
-)
-
-from codecritic.utils.json import load_json
-from codecritic.dataset.utils import save_jsonl_dataset
-from codecritic.utils.vllm import vllm_chatcomplete
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model", type=str)
-    parser.add_argument("--preference_dataset", type=str)
-    parser.add_argument("--output_dir", type=str)
-    args = parser.parse_args()
-
-    preference_dataset = load_json(args.preference_dataset)
-    cov_prompts = [convert_preference_to_vot_prompt(x) for x in preference_dataset]
-    cov_prompts = list(chain(*cov_prompts))
-
-    sampling_params = dict(n=1, temperature=0.0, max_tokens=2048)
-    covs = vllm_chatcomplete(args.model, cov_prompts, sampling_params)
-    dataset = list(map(convert_cov_to_cov_dataset, covs))
-
-    save_jsonl_dataset(dataset, args.output_dir)

codecritic/cli/test_genrm.py

@@ -5,8 +5,8 @@ import os
 from transformers import AutoTokenizer
 from vllm import SamplingParams
 
-from codecritic.dataset.genrm_prompt import THINK_PROMPT, JUDGE_PROMPT, JUDGE_TOEKNS
-from codecritic.evaluation.inference import generate_worker, score_worker
+from codecritic.dataset.genrm_prompt import THINK_MESSAGE, 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
 
@@ -35,7 +35,7 @@ if __name__ == "__main__":
 
     if args.reasoning:
         for item in dataset:
-            item["messages"].append({"role": "user", "content": THINK_PROMPT})
+            item["messages"].append(THINK_MESSAGE)
 
         sampling_params = SamplingParams(
             n=1,
@@ -58,7 +58,7 @@ if __name__ == "__main__":
     negative_token = get_token_id(JUDGE_TOEKNS["negative"])
 
     for item in dataset:
-        item["messages"].append({"role": "user", "content": JUDGE_PROMPT})
+        item["messages"].append(JUDGE_MESSAGE)
 
     worker = partial(
         score_worker,

codecritic/dataset/genrm_prompt.py

 JUDGE_PROMPT = "Is the code correct (Yes/No)?"
+JUDGE_MESSAGE = {"role": "user", "content": JUDGE_PROMPT}
 JUDGE_TOEKNS = {
     "positive": "Yes",
     "negative": "No"
 }
 
-def mk_critic_verify(answer=None):
-    # answer: bool or none
-    message = [{"role": "user", "content": JUDGE_PROMPT}]
-    if answer is not None:
-        response = "Yes" if answer else "No"
-        message.append({"role": "assistant", "content": response})
-
-    return message
-
+def mk_judge_response(response):
+    "response: positive or negative"
+    return {"role": "assistant", "content": JUDGE_TOEKNS[response]}
 
+THINK_PROMPT = ""
+THINK_MESSAGE = {"role": "user", "content": THINK_PROMPT}
 

codecritic/dataset/reformat.py

+import codecritic.dataset.genrm_prompt as genrm_prompt
+
+
+def reward_to_sft(item):
+    messages = item.pop("messages")
+    chosen_response = item.pop("chosen")
+    rejected_response = item.pop("rejected")
+
+    chosen = {
+        **item,
+        "question": messages + chosen_response + [genrm_prompt.JUDGE_MESSAGE],
+        "response": genrm_prompt.mk_judge_response("positive"),
+    }
+    rejected = {
+        **item,
+        "question": messages + rejected_response + [genrm_prompt.JUDGE_MESSAGE],
+        "response": genrm_prompt.mk_judge_response("negative"),
+    }
+    return chosen, rejected

codecritic/evaluation/reformat.py

+"""
+changing the format of the dataset.
+"""
\ No newline at end of file

codecritic/evaluation/inference.py → codecritic/utils/inference.py

@@ -51,26 +51,6 @@ def generate_worker(cuda_device, prompts, model_path, sampling_params):
 
 
 def score_worker(cuda_device, prompts, model_path, positive_token, negative_token=None):
-
-    def only_positive(logprob):
-        positive_logprob = logprob.get(positive_token)
-        positive_prob = np.exp(positive_logprob.logprob) if positive_logprob else 0
-        return {"score": positive_prob}
-
-    def pos_and_neg(logprob):
-        positive_logprob = logprob.get(positive_token)
-        positive_prob = np.exp(positive_logprob.logprob) if positive_logprob else 0
-
-        negative_logprob = logprob.get(negative_token)
-        negative_prob = np.exp(negative_logprob.logprob) if negative_logprob else 0
-
-        return {
-            "score": positive_prob / (positive_prob + negative_prob),
-            "uncertainty": 1 - (positive_prob + negative_prob),
-        }
-
-    compute_score = only_positive if negative_token is None else pos_and_neg
-
     os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(cuda_device)
 
     llm = LLM(
@@ -97,9 +77,21 @@ def score_worker(cuda_device, prompts, model_path, positive_token, negative_toke
         assert len(output.outputs) == 1, "The scorer must provide a single score."
         for response in output.outputs:
             # response.logprobs: list[dict[int, Logprob]] https://github.com/vllm-project/vllm/blob/main/vllm/sequence.py
-            scores = compute_score(response.logprobs[0])
+            logprob = response.logprobs[0]
+            positive_logprob = logprob.get(positive_token)
+            positive_prob = np.exp(positive_logprob.logprob) if positive_logprob else 0
+
+            if negative_token:
+                negative_logprob = logprob.get(negative_token)
+                negative_prob = np.exp(negative_logprob.logprob) if negative_logprob else 0
+            else:
+                negative_prob = None
             text = response.text
 
-            results.append({**item, **scores, "critic_text": text})
+            result = item.copy()
+            result["positive_score"] = positive_prob
+            result["negative_score"] = negative_prob
+            result["meta_score_response"] = text
+            results.append(result)
 
     return results