Upload 2 files

special_token_test.py

@@ -0,0 +1,37 @@
+from transformers import AutoTokenizer
+import os
+
+# 模型路徑
+MODEL_PATH = "/home/at0842/ycl466704.ai13/.cache/huggingface/hub/models--openai--gpt-oss-20b/snapshots/6cee5e81ee83917806bbde320786a8fb61efebee"
+
+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 Tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
+
+# 1. 取得官方定義的特殊 ID 列表
+# 注意：在 Transformers 中，通常是使用 all_special_ids
+special_ids = tokenizer.all_special_ids
+
+print("\n" + "="*50)
+print(f"📊 官方定義的特殊 ID 列表 (共 {len(special_ids)} 個):")
+print(special_ids)
+print("="*50)
+
+# 2. 解碼這些 ID 看看它們分別是什麼標籤
+print("\n🔍 特殊 ID 與對應文字明細：")
+for tid in sorted(special_ids):
+    t_text = tokenizer.decode([tid])
+    # 印出 ID 和對應的文字，repr 幫助看清空格或換行
+    print(f"ID: {tid:7} | Text: {repr(t_text):15}")
+
+# 3. 額外檢查：看看是否有大於 200,000 的 Added Tokens 沒被列入
+# 有些自定義頻道標籤可能在 added_tokens 裡但不在 all_special_ids
+print("\n" + "="*50)
+print("💡 額外檢查：Added Tokens (自定義標籤)")
+added_tokens = tokenizer.get_added_vocab()
+for token_str, token_id in sorted(added_tokens.items(), key=lambda x: x[1]):
+    if token_id not in special_ids:
+        print(f"ID: {token_id:7} | Text: {repr(token_str):15} (不在官方 special_ids 內)")

to_txt.py

@@ -0,0 +1,49 @@
+import json
+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 = "all_dupcleaned_data_turn.jsonl"
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
+
+inspect_logs = []
+
+print(f"⏳ 開始生成診斷檔案...")
+
+with open(INPUT_JSONL, "r", encoding="utf-8") as f_in:
+    for i, line in enumerate(f_in):
+        if i >= 10: break # 只看前 10 筆
+        
+        try:
+            entry = json.loads(line)
+            messages = entry.get("messages", [])
+            
+            # 1. 取得完整序列 (模擬訓練時的狀態)
+            full_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=False)
+            
+            # 2. 取得 Prompt 結尾位置 (模擬訓練時的 Mask 起點)
+            context_ids = tokenizer.apply_chat_template(messages[:-1], tokenize=True, add_generation_prompt=True)
+            start_idx = len(context_ids)
+
+            inspect_logs.append(f"=== CASE {i+1} ===\n")
+            inspect_logs.append(f"Prompt 結束位置 (start_idx): {start_idx}\n")
+            inspect_logs.append(f"Assistant 區段 Token 明細 (從 Index {start_idx} 開始):\n")
+            
+            # 3. 逐個印出 Token ID 和解碼文字，看看標籤長在哪
+            # 我們多印 20 個，保證看到真正的內容
+            end_range = min(start_idx + 20, len(full_ids))
+            for idx in range(start_idx, end_range):
+                tid = full_ids[idx]
+                t_text = tokenizer.decode([tid])
+                # 修正後的屬性名稱：all_special_ids
+                is_special = "[SPECIAL]" if tid in tokenizer.all_special_ids else ""
+                inspect_logs.append(f"Index {idx:4} | ID: {tid:7} | Text: {repr(t_text):15} {is_special}\n")
+            
+            inspect_logs.append("-" * 60 + "\n")
+        except Exception as e:
+            print(f"跳過第 {i} 筆，原因: {e}")
+
+with open("inspect_tokens.txt", "w", encoding="utf-8") as f:
+    f.writelines(inspect_logs)
+
+print("✅ 診斷檔案已生成: inspect_tokens.txt")