Problem
convert_to_sharegpt.py converts every sample into a single [user, assistant] pair. Real conversations with back-and-forth exchanges are never represented as multi-turn samples in the ShareGPT format.
This significantly limits fine-tuning quality because the model never learns to continue a thread — it only learns isolated single-shot responses.
Expected behavior
The pipeline should support grouping related exchanges into multi-turn ShareGPT conversations:
{
"conversations": [
{"from": "user", "value": "what did you think of the movie?"},
{"from": "assistant", "value": "loved it"},
{"from": "user", "value": "which part?"},
{"from": "assistant", "value": "the ending"}
]
}This depends on issue #3 (conversation context grouping) being resolved first.
Problem
convert_to_sharegpt.pyconverts every sample into a single[user, assistant]pair. Real conversations with back-and-forth exchanges are never represented as multi-turn samples in the ShareGPT format.This significantly limits fine-tuning quality because the model never learns to continue a thread — it only learns isolated single-shot responses.
Expected behavior
The pipeline should support grouping related exchanges into multi-turn ShareGPT conversations:
{ "conversations": [ {"from": "user", "value": "what did you think of the movie?"}, {"from": "assistant", "value": "loved it"}, {"from": "user", "value": "which part?"}, {"from": "assistant", "value": "the ending"} ] }This depends on issue #3 (conversation context grouping) being resolved first.