step4: eval scored results

example_config.toml

@@ -62,6 +62,7 @@ deepspeed_cfg_path = ""
 [critic.test]
 reason_result_path = ""
 score_result_path = ""
+eval_result_path = ""
 
 [critic.test.sampling_params]
 n = 1
@@ -80,4 +81,5 @@ deepspeed_cfg_path = ""
 
 [sftorm.test]
 prompt_path = ""
-score_result_path = ""
+score_result_path = ""
\ No newline at end of file
+eval_result_path = ""
\ No newline at end of file

step4_test_reward_model_client.py

@@ -18,6 +18,6 @@ if __name__ == "__main__":
     # results = load_jsonl(result_path)
 
     groups = group_results(results, cfg["apps"])
-    eval_results = [score_pass_at_k(groups, k, orm_test_model) for k in range(1, 32)]
+    eval_results = [score_pass_at_k(groups, k, orm_test_model) for k in range(1, 16)]
     save_jsonl(eval_results, cfg["orm"][orm_test_model]["eval_result_path"])
     print(eval_results)

step4_test_sftorm.py

 from utils_vllm import vllm_score
 from utils import read_config, load_jsonl, save_jsonl, extract_code
-from utils_preference_dataset import code_template, mk_critic_qa, mk_critic_verify, mk_sft_item
+from utils_preference_dataset import code_template, mk_critic_qa, mk_critic_verify
+from utils_metric import group_results, score_pass_at_k
 from transformers import AutoTokenizer
 
 
@@ -29,4 +30,10 @@ if __name__ == "__main__":
         cfg["sftorm"]["test"]["prompt_path"],
         cfg["sftorm"]["test"]["score_result_path"],
         score_token
-    )
+    )
\ No newline at end of file
+
+    results = load_jsonl(cfg["sftorm"]["test"]["score_result_path"])
+    groups = group_results(results, cfg["apps"])
+    eval_results = [score_pass_at_k(groups, k, "sft-orm") for k in range(1, 16)]
+    save_jsonl(eval_results, cfg["sftorm"]["test"]["eval_result_path"])
+    print(eval_results)
\ No newline at end of file