Finetuning

Finetuning#

This page documents the experimental fine-tuning support that still exists in the repository. It is not a maintained end-to-end training guide.

Current scope#

Today gptme ships one helper for this workflow: scripts/train/collect.py.

That script:

  • reads local conversation logs from ~/.local/share/gptme/logs/

  • filters out generated-name/test conversations and low-quality chats

  • strips leading system prompts

  • renders each conversation with the selected Hugging Face chat template

  • writes train.csv and train.jsonl

The files are written to your current working directory, not to a train/ subdirectory in the repo.

Collect local conversations#

From the repository root, run:

./scripts/train/collect.py --model "HuggingFaceH4/zephyr-7b-beta"

Pick a model whose chat template matches the format you want in the resulting training data. The script uses transformers.pipeline(...) and tokenizer.apply_chat_template(...) to turn stored chats into prompt text. Run it in an environment where torch and transformers are already installed.

What is not maintained here#

This page intentionally does not claim a supported workflow for:

  • mixing in OpenAssistant or other public datasets

  • importing exported ChatGPT conversations

  • splitting train and validation sets

  • training with a specific stack such as Axolotl, Transformers, or OpenPipe

Older revisions of this page sketched those steps, but they were incomplete and had drifted out of sync with the repo. Until gptme grows a maintained fine-tuning pipeline again, treat this document as a narrow note about the collector script rather than a full recipe.

Training stack pointers#

If you want to take the exported data further, start with upstream docs for the training stack you actually plan to use:

Model suggestions#

  • HuggingFaceH4/zephyr-7b-beta

  • teknium/Replit-v2-CodeInstruct-3B

    • This was previously used for testing/debugging, but availability and local hardware support may vary.