agentscope-ai / agentscope-ai/Trinity-RFT

[FEATURE] Recommended way of producing multiple Experiences in multi-turn dialogs

未关闭
#269 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
enhancement
主要语言
Python
星标
701
派生
79
平均合并
8 小时 7 分钟
30 天内合并 PR
1

描述

Imagine a two-turn dialog:
1. system prompt
2. first user prompt
3. first assistant response
4. second user prompt
5. final assistant response

One could represent it as Experience by concatenating everything and masking everything except two assistant responses

Another way could be splitting it into two experiences.
The first one:
1. system prompt
2. first user prompt
3. first assistant response (masked as target assistant)

The second one:
1. system prompt
2. first user prompt
3. first assistant response
4. second user prompt
5. final assistant response (masked as target assistant)

Is this other multi-experience way recommended? If so, how to best produce it?

E.g. should a `run()` method return these several experiences, right?

贡献指南

打开贡献指南

评估

这个 Issue 还没有评估数据。

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。