Skip to content

System Prompt and Format Reward inconsistent in stage 1 #12

@RayquazaX

Description

@RayquazaX

1.1 thinking vs think
Parquet 训练数据(data/bfcl_train_base.parquet、data/bfcl_val.parquet)中全部 200 条 system prompt 使用 thinking /thinking 作为 XML 标签,但以下两处代码只匹配 think /think:

  • env_tuning/interaction/utils.py:46 — parse_model_response() 正则 think ... /think, 不匹配 thinking
  • env_tuning/interaction/execution_manager.py:110 — format_execution_response() 中 user_hint 使用 think /think

影响:模型按 system prompt 输出 thinking,parser 只匹配 think,所有输出在第一步即被判定格式错误(score = -3)

1.2 answer 标签未在 system prompt 中定义
System prompt 对不需要工具调用的情况指示为:

  • "If no tool calls are necessary or possible: Directly provide a user-facing response in plain text."

但 env_tuning/interaction/utils.py:85-89 的 parse_model_response() 要求必须用 answer.../answer 包裹最终回答,否则返回格式错误。answer 仅在env_tuning/interaction/execution_manager.py:110-112 的 user_hint 中后续引入,模型在初始 turn 无从得知需要使用该标签。

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions