Wtf. I don't know what happened. Reason-ORM seems a perfect judger. I must save…

Wtf. I don't know what happened. Reason-ORM seems a perfect judger. I must save this version of code

codecritic/cli/convert_pairwise_to_sft.py

@@ -10,20 +10,20 @@ from codecritic.data.verify import mk_critic_verify
 
 
 def mk_sft_dataset(messages):
-    assert len(messages) == 4
-    question = messages[:3]
-    response = messages[3:]
+    question = messages[:-1]
+    response = messages[-1:]
     return dict(question=question, response=response)
 
 
 def convert_preference_to_sft(item):
-    message = item["messages"][0]["content"]
-    chosen = item["chosen"][0]["content"]
-    rejected = item["rejected"][0]["content"]
+    message = item["messages"]
+    chosen = item["chosen"]
+    rejected = item["rejected"]
 
 
-    messages1 = mk_message(message, chosen) + mk_critic_verify(True)
-    messages2 = mk_message(message, rejected) + mk_critic_verify(False)
+
+    messages1 = message + chosen + mk_critic_verify(True)
+    messages2 = message + rejected + mk_critic_verify(False)
     return mk_sft_dataset(messages1), mk_sft_dataset(messages2)
 
 

codecritic/cli/run_rm_test.py

@@ -22,12 +22,7 @@ def get_rewards_from_server(server_url: str, messages: list[str]):
     return rewards
 
 
-def test_reward_model(server_url, item, tokenizer, preprocess=True):
-    if preprocess:
-        response = item["messages"][-1]["content"]
-        code = code_template.format(extract_code(response))
-        item["messages"][-1]["content"] = code
-
+def test_reward_model(server_url, item, tokenizer):
     query = tokenizer.apply_chat_template(item["messages"], tokenize=False)
     score = get_rewards_from_server(server_url, [query])[0]
 
@@ -44,7 +39,6 @@ if __name__ == "__main__":
     parser.add_argument("--model", type=str)
     parser.add_argument("--test", type=str)
     parser.add_argument("--apps", type=str)
-    parser.add_argument("--skip_preprocess", action="store_false")
     args = parser.parse_args()
 
     home_path = Path(args.model).parent
@@ -55,7 +49,7 @@ if __name__ == "__main__":
     test_dataset = load_jsonl(args.test)
     server_url = "http://0.0.0.0:5000/get_reward"
     tokenizer = AutoTokenizer.from_pretrained(args.model)
-    results = [test_reward_model(server_url, item, tokenizer, args.skip_proprocess) for item in tqdm(test_dataset)]
+    results = [test_reward_model(server_url, item, tokenizer) for item in tqdm(test_dataset)]
     score_path = result_dir / "scores.jsonl"
     save_jsonl(results, score_path)