#2424th of 53 in Training and fine-tuning
skillopt
Trains agent skill documents against a frozen model with validation-gated edits
- Stars
- 19k
- License
- MIT
- Last commit
- Nov 2026
- Last release
- Jul 2026
- Language
- Python
7 on 5 May 202618,992 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
SkillOpt treats a single skill document as the trainable state of a frozen agent. A separate optimizer model turns scored rollouts into bounded add/delete/replace edits, and in the default path an edit is kept only if it strictly improves a held-out validation score. The output is a compact best_skill.md (typically 300-2,000 tokens) with no extra inference-time model calls. It installs from PyPI and includes a nightly SkillOpt-Sleep CLI and an optional Gradio dashboard.
Who it is for: Engineers tuning agent skill files for Codex, Claude Code or chat models
Strengths
- Validation-gated edits with a textual learning-rate budget and rejected-edit buffer
- Deployed skill is one markdown file, with no added inference-time model calls
- Backends for OpenAI, Azure, Claude, Qwen, MiniMax, plus any OpenAI-compatible endpoint
- Also drives Codex CLI and Claude Code harnesses; six built-in benchmarks
Weaknesses
- Training needs a separate optimizer model and scored rollouts, so API cost scales with epochs
- Claude Code, Codex, Copilot and Devin integration files are in the repo, not the PyPI wheel
- Main CLI does not expose every experimental control (multi-objective, replay, dream rollouts) as flags
- Reported gains come from the authors' own benchmarks; no independent reproduction is cited
What it needs
- no GPU
- Needs Python 3.10+, LLM API access (OpenAI, Azure, Claude, Qwen, MiniMax or OpenAI-compatible)
- Models: OpenAI, Azure OpenAI, Claude, Qwen, MiniMax
- port 7860
Quick answers
Is skillopt open source?
Yes. Its code is on GitHub under the MIT license.
Does skillopt need a GPU?
No, it runs without one.
Also in Training and fine-tuning
See all 53| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 88 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 86 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 86 out of 100 |