#统一方案v2（v1基础上统一截断、填充、eos_token等） #数据读取改为json格式，适配chat模型 #定义新的格式化函数：利用 chat_template 将列表转为字符串 #添加 config.pad_token_id = tokenizer.pad_token_id #将 config.std 改为 getattr(config, "std", None)，解决配置中缺失 mean/std 字段时的 AttributeError #将 batch["chosen_prompt"] 替换为 batch["messages"]，将 batch["reject"] 替换为 batch["rejected"]，适配 JSON 数据格式 #在 sample_table.add_data 中对列表数据（如 messages）添加 str() 转换，防止日志记录出错 import torch, wandb, pandas as pd from tqdm import tqdm from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig import os # === 参数 === rm_path = "/home/LLaMA-Factory/output/mistral-7B_rm_instag_281_1_64_A_6k*3_BCD_m/merged_model" data_path = "/home/DataProcess/data/test_1w_chatml_closed_numdel_replaced.json" save_path = "/home/DataProcess/data/test_1w_chatml_closed_numdel_replaced_scored_by_rmv3.json" batch_size = 16 max_length = 4096 # === wandb === wandb.init(project="reward_model_scoring", name="rmv2.2_acc_1") # === 模型 & tokenizer === tokenizer = AutoTokenizer.from_pretrained(rm_path, trust_remote_code=True) tokenizer.padding_side = "left" # ← 修改 if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token tokenizer.pad_token_id = tokenizer.eos_token_id config = AutoConfig.from_pretrained(rm_path) config.num_labels = 1 # reward head config.pad_token_id = tokenizer.pad_token_id model = AutoModelForSequenceClassification.from_pretrained( rm_path, config=config, device_map="auto") model.config.pad_token_id = tokenizer.pad_token_id model.eval() device = next(model.parameters()).device # 读取 json 文件 df = pd.read_json(data_path).reset_index(drop=True) # 定义新的格式化函数：利用 chat_template 将列表转为字符串 def format_chat_input(history, response_list, tokenizer): """ history: list, 对应 json 中的 "messages" response_list: list, 对应 json 中的 "chosen" 或 "rejected" """ # 拼接历史对话和当前的回复 full_conversation = history + response_list # 使用 tokenizer 的聊天模板转为字符串 (例如转为 <|im_start|>user...<|im_end|>) # tokenize=False 表示只返回字符串，不返回 ID try: txt = tokenizer.apply_chat_template(full_conversation, tokenize=False, add_generation_prompt=False) except Exception as e: # 如果模型没有模板（很少见），则手动拼接作为兜底 txt = "" for msg in full_conversation: txt += f"{msg['role']}: {msg['content']}\n" # 依然保留 V2 脚本的核心逻辑：强制检查并添加 EOS if not txt.endswith(tokenizer.eos_token): txt += tokenizer.eos_token return txt def encode_batch(chosen_texts, rejected_texts, tokenizer, max_length, device): # 1 tokenize ch = tokenizer(chosen_texts, add_special_tokens=False, truncation=True, max_length=max_length, padding=False) rj = tokenizer(rejected_texts, add_special_tokens=False, truncation=True, max_length=max_length, padding=False) ids1, mask1 = ch["input_ids"], ch["attention_mask"] ids2, mask2 = rj["input_ids"], rj["attention_mask"] # 2 ensure eos 存在 for arr_ids, arr_mask in ((ids1, mask1), (ids2, mask2)): for i in range(len(arr_ids)): arr_ids[i][-1] = tokenizer.eos_token_id arr_mask[i][-1] = 1 # 3 left-pad 到 joint_max joint_max = max(max(len(x) for x in ids1), max(len(x) for x in ids2)) lpad = lambda seq, pad: [pad]*(joint_max-len(seq)) + seq ids1 = [lpad(x, tokenizer.pad_token_id) for x in ids1] ids2 = [lpad(x, tokenizer.pad_token_id) for x in ids2] mask1 = [lpad(x, 0) for x in mask1] mask2 = [lpad(x, 0) for x in mask2] input_ids = torch.tensor(ids1 + ids2, dtype=torch.long).to(device) attn_masks = torch.tensor(mask1 + mask2, dtype=torch.long).to(device) return input_ids, attn_masks, len(chosen_texts) # === 推理 === chosen_scores, rejected_scores, accs = [], [], [] sample_table = wandb.Table(columns=["index","prompt","chosen","rejected", "chosen_score","rejected_score","delta","acc"]) for i in tqdm(range(0, len(df), batch_size)): batch = df.iloc[i:i+batch_size] chosen_texts = [] rejected_texts = [] # 必须遍历每一行，因为 apply_chat_template 不能直接处理 DataFrame 列 for _, row in batch.iterrows(): # row["messages"] 是历史记录 # row["chosen"] 是 [{"role": "assistant", "content": "..."}] c_txt = format_chat_input(row["messages"], row["chosen"], tokenizer) r_txt = format_chat_input(row["messages"], row["rejected"], tokenizer) chosen_texts.append(c_txt) rejected_texts.append(r_txt) input_ids, attn_masks, split = encode_batch(chosen_texts, rejected_texts, tokenizer, max_length, device) with torch.no_grad(): rewards = model(input_ids=input_ids, attention_mask=attn_masks).logits.squeeze(-1) model_std = getattr(config, "std", None) model_mean = getattr(config, "mean", None) if model_std is not None and model_mean is not None: rewards = rewards * model_std + model_mean chosen_r, rejected_r = rewards[:split], rewards[split:] for j in range(len(chosen_r)): idx = i + j c, r = chosen_r[j].item(), rejected_r[j].item() delta = c - r acc = int(delta > 0) chosen_scores.append(c) rejected_scores.append(r) accs.append(acc) avg_acc = sum(accs) / len(accs) print(f"[{idx}] acc={acc}, chosen={c:.3f}, rejected={r:.3f}, Δ={delta:.3f} | avg acc={avg_acc:.3f}") sample_table.add_data(idx, str(batch["messages"].iloc[j]), # Prompt (历史对话) str(batch["chosen"].iloc[j]), # Chosen str(batch["rejected"].iloc[j]), # Rejected c, r, delta, acc) # === 结果 === df["chosen_score"] = chosen_scores df["rejected_score"] = rejected_scores df["delta"] = df["chosen_score"] - df["rejected_score"] df["acc"] = accs accuracy = df["acc"].mean() mean_chosen = df["chosen_score"].mean() mean_reject = df["rejected_score"].mean() mean_delta = df["delta"].mean() print(f"\nAccuracy = {accuracy:.3f}") print(f"mean_chosen = {mean_chosen:.3f}, mean_rejected = {mean_reject:.3f}, mean_delta = {mean_delta:.3f}") # 确保目录存在 os.makedirs(os.path.dirname(save_path), exist_ok=True) # 保存为 JSON 列表格式 (orient="records")，确保中文不转义 (force_ascii=False) df.to_json(save_path, orient="records", force_ascii=False, indent=4) print(f"\n[Success] Scored data saved to: {save_path}") wandb.log({ "samples_table": sample_table, "final_accuracy": accuracy, "mean_chosen_score": mean_chosen, "mean_rejected_score": mean_reject, "mean_delta_score": mean_delta, }) wandb.finish()