trans4.py

import json
import os
from transformers import AutoTokenizer

# 1. 配置
MODEL_PATH = "/home/at0842/ycl466704.ai13/.cache/huggingface/hub/models--openai--gpt-oss-20b/snapshots/6cee5e81ee83917806bbde320786a8fb61efebee"
INPUT_JSONL = "aug_data_random_para_turn.jsonl"
OUTPUT_JSONL = "aug_data_random_para_turn_gpt_oss_20b_pretokenized.jsonl"
DEBUG_FILE = "debug_readable_aug_data_random_para_turn.txt"
MAX_LENGTH = 14436 

if not os.path.exists(MODEL_PATH):
    print(f"❌ 錯誤：找不到模型路徑 {MODEL_PATH}")
    exit()

print(f"⏳ 正在初始化 GPT-OSS 20B Tokenizer...")
# fix_mistral_regex=True
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
RETURN_TOKEN_ID = tokenizer.convert_tokens_to_ids("<|return|>")

def process_entry(entry, index):
    try:
        messages = entry.get("messages", [])
        raw_tools = entry.get("tools", None) 
        
        # 工具格式化
        formatted_tools = None
        if raw_tools and isinstance(raw_tools, list):
            formatted_tools = [{"type": "function", "function": t} if not (isinstance(t, dict) and "function" in t) else t for t in raw_tools]

        if not isinstance(messages, list) or len(messages) < 2:
            return None, None, "對話輪次不足"
        
        # 1. 取得序列
        full_ids = tokenizer.apply_chat_template(messages, tools=formatted_tools, tokenize=True, add_generation_prompt=False, truncation=False)
        if full_ids[-1] != RETURN_TOKEN_ID:
            full_ids.append(RETURN_TOKEN_ID)
            
        context_ids = tokenizer.apply_chat_template(messages[:-1], tools=formatted_tools, tokenize=True, add_generation_prompt=True, truncation=False)   
        start_idx = len(context_ids)
        actual_length = len(full_ids)

        # 2. 舊版對照
        old_target_tokens = full_ids[start_idx:actual_length]
        old_target_text = tokenizer.decode(old_target_tokens)

        # 3. 雙區間精確過濾邏輯
        new_labels = [-100] * actual_length
        CHANNEL_ID = 200005 
        MESSAGE_ID = 200008 
        SPECIAL_BOUNDARY = 199998 # 統一使用你測出的 199998
        
        # 找出第一個 <|channel|>
        first_channel_idx = -1
        for i in range(start_idx, actual_length):
            if full_ids[i] == CHANNEL_ID:
                first_channel_idx = i
                break
        
        # 處理區間 A：指令區 (只有在找到 CHANNEL_ID 時才執行)
        if first_channel_idx != -1:
            for i in range(start_idx, first_channel_idx):
                tid = full_ids[i]
                if tid < SPECIAL_BOUNDARY and tid not in tokenizer.all_special_ids:
                    new_labels[i] = tid

        # 處理區間 B：保留 <|channel|> 到 <|message|> 之間的非 special token
        # 例如：commentary json / final
        is_in_channel_zone = False
        is_in_message_zone = False

        for i in range(start_idx, actual_length):
            tid = full_ids[i]

            # 進入 channel 區
            if tid == CHANNEL_ID:
                is_in_channel_zone = True
                is_in_message_zone = False
                continue

            # 進入 message 區
            if tid == MESSAGE_ID:
                is_in_channel_zone = False
                is_in_message_zone = True
                continue

            # 在 channel 區：只保留非 special token
            if is_in_channel_zone:
                if tid < SPECIAL_BOUNDARY and tid not in tokenizer.all_special_ids:
                    new_labels[i] = tid
                else:
                    is_in_channel_zone = False

            # 在 message 區：保留內容，遇到下一個 special token 就結束
            elif is_in_message_zone:
                if tid < SPECIAL_BOUNDARY and tid not in tokenizer.all_special_ids:
                    new_labels[i] = tid
                else:
                    is_in_message_zone = False
        # 4. 生成新版 Target 文字 (用於 debug)
        valid_new_tokens = [full_ids[x] for x in range(start_idx, actual_length) if new_labels[x] != -100]
        new_target_text = tokenizer.decode(valid_new_tokens)

        # 建立可視化結果
        visual_str = ""
        for i in range(actual_length):
            t_text = tokenizer.decode([full_ids[i]])
            if new_labels[i] == -100:
                visual_str += f"[MASK:{t_text}]"
            else:
                visual_str += f"**{t_text}**"

        result = {"input_ids": full_ids, "attention_mask": [1] * actual_length, "labels": new_labels}
        comparison = {"old": old_target_text, "new": new_target_text, "visual": visual_str}

        return result, comparison, None
    except Exception as e:
        return None, None, str(e)

# --- 執行主程式 ---
success_count = 0
comparison_logs = []

print(f"  開始處理資料並生成對照檢查檔...")

with open(INPUT_JSONL, "r", encoding="utf-8") as f_in, \
     open(OUTPUT_JSONL, "w", encoding="utf-8") as f_out:
    
    for i, line in enumerate(f_in):
        line = line.strip()
        if not line: continue
        
        try:
            entry = json.loads(line)
            result, comp, err = process_entry(entry, i)
            
            if result:
                # 修正：寫入時建議確保 ensure_ascii=False，否則中文會變編碼
                f_out.write(json.dumps(result, ensure_ascii=False) + "\n")
                
                if success_count < 20: # 稍微多存一點抽查
                    comparison_logs.append(f"### CASE {success_count+1} (Index {i}) ###\n")
                    comparison_logs.append(f"【舊版 Target】:\n{comp['old']}\n")
                    comparison_logs.append(f"【新版 Target】:\n{comp['new']}\n")
                    comparison_logs.append(f"【可視化預覽】:\n{comp['visual']}\n")
                    comparison_logs.append("-" * 100 + "\n\n")
                success_count += 1
            else:
                if i < 100: # 只印前 100 筆的錯誤，避免洗板
                    print(f" 跳過第 {i} 筆: {err}")
        except Exception as main_e:
            print(f" 第 {i} 筆發生嚴重錯誤: {main_e}")

with open(DEBUG_FILE, "w", encoding="utf-8") as f_debug:
    f_debug.writelines(comparison_logs)

print(f"  處理完成！")
print(f" 1. 訓練用 Token 檔: {OUTPUT_JSONL}")
print(f" 2. 人類可讀對照檔: {DEBUG_FILE}")
print(f" 總共成功處理: {success_count} 筆")



        # # 處理區間 B：內容區 (掃描全域)
        # is_in_message_zone = False
        # for i in range(start_idx, actual_length):
        #     tid = full_ids[i]
        #     if tid == MESSAGE_ID:
        #         is_in_message_zone = True
        #         continue 
            
        #     # 遇到下一個標籤就關閉
        #     if is_in_message_zone and (tid >= SPECIAL_BOUNDARY or tid in tokenizer.all_special_ids):
        #         is_in_message_zone = False
            
        #     if is_in_message_zone:
        #         new_labels[i] = tid