All the proxy logs I could find (lmk if there are more), converted to ShareGPT so it's all formatted the same way in one dataset. I didn't .strip() any of the turns, I only used ftfy.fix_text() on them, and skipped any empty turns.
The original sample response is moved to be the final turn of conversations. If the last turn in the original sample prompt was a model turn, it is assumed that it was for doing prefill. So I moved this turn to response_prefill.
sample["response_prefill"] and… See the full description on the dataset page:
https://huggingface.co/datasets/PJMixers-Dev/proxy-logs-ShareGPT.
