訓練資料
single turn:

source:
1. squad
2. tri

attack:
1. none
2. naive
3. ignore
4. escape(eval only)
5. completion(eval only)
6. conv_attack(eval only)

attack位置:
1. 左邊
2. 右邊

configuration:
1. Prompt-Based Separator
2. Native Tool Separator(Without tool query, only tool response)
3. Native Tool Separator(Empty tool query)

測試資料:
messages生成沿用訓練資料
補上 defense 串 LLm 生成結果
加上判斷結果