llm-finetuning

PackAI & models

llm-finetuning is a plugin that lets an AI agent run the full fine-tuning lifecycle for open-weight language models. It covers building an evaluation harness first, choosing a method and model, preparing data, training with Unsloth or TRL, gating checkpoints, and exporting GGUF or FP files. The first keyword phrase, eval-gated LLM fine-tuning lifecycle, describes its core flow: checkpoints only pass when evaluations approve them.

Unavailable. Delivery for this kind is on the roadmap — not serving yet.

Have an open-weights language model you want to fine-tune.

What your AI can do with it

  • Builds an eval harness before training starts
  • Selects fine-tuning method and model (LoRA/QLoRA, DPO, GRPO/RLVR, vision SFT)
  • Prepares training data
  • Trains with Unsloth or TRL
  • Gates checkpoints on evaluation results
  • Exports quantized GGUF/FP files

Getting started

  1. Have an open-weights language model you want to fine-tune.
  2. Add the llm-finetuning plugin to your agent setup.
  3. Ask the agent to build the eval harness first.
  4. Have the agent select the method and model, prepare data, and run training with Unsloth or TRL.
  5. Review gated checkpoints and export the result as GGUF or FP.

Signals

GitHub stars
40k
Forks
4k
Last commit
Sep 2026

Questions

What training methods does it support?
LoRA and QLoRA, DPO, GRPO/RLVR, and vision SFT, with quantized export of the trained model.
Which training libraries does it use?
Training runs with Unsloth or TRL.
What export formats are available?
It exports GGUF and FP files, including quantized exports.
Why build the eval harness first?
The lifecycle is eval-gated: checkpoints are gated on evaluation results, so the harness is set up before training begins.
Advanced
Item type
plugin
Key
wshobson-agents-llm-finetuning
Source
github.com/wshobson/agents