How should I structure data for fine-tuning a model in a groupchat setting?

[PyroGenesis]

I'm trying to finetune some llm agents such that they perform some specialized task. However, I am unsure of how to best convert an Autogen groupchat conversation into a training format.

For example, let's say I'm working with the following conversation:

User prompt
Agent 1 speaks
Agent 2 speaks
Agent 3 speaks

Now let's say I'm trying to fine-tune Agent 3 here, which is backed by a Llama 2 model

Llama 2 model format:

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ user_msg_1 }} [/INST] {{ model_answer_1 }} </s><s>[INST] {{ user_msg_2 }} [/INST]

How should I go about structuring this into training data? The 3 ways I've thought about are:

Option 1: Treat other Agent responses as separate user responses

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ User prompt }} [/INST] {{ Agent 1 output }} [/INST] {{ Agent 2 output }} [/INST] </s><s>[INST] {{ Agent 3 output }} [/INST]

Option 2: Combine other Agent responses into single User response

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ User prompt }} {{ Agent 1 output }} {{ Agent 2 output }} [/INST] </s><s>[INST] {{ Agent 3 output }} [/INST]

Option 3: Treat other Agent responses as AI responses

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ User prompt }} [/INST] </s><s>[INST] {{ Agent 1 output }} [/INST] [INST] {{ Agent 2 output }} [/INST] [INST] {{ Agent 3 output }} [/INST]

Do any of these look viable? Or should I try and roll my own new format which would be more compatible with groupchat conversations?

[rickyloynd-microsoft]

@afourney fyi

[amanAtHoneyHealth]

I am also looking for an answer to this.