菜单
Use this project skill when Michael asks to fine-tune for OpenClaw, run the pipeline, build training datasets, run GRPO, create LoRA adapters, extract Claude sessions, convert Anthropic to OpenAI format, enrich OpenClaw tools, or work in the OpenClaw pipeline directory.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | openclaw-finetuning |
| description | Use this project skill when Michael asks to fine-tune for OpenClaw, run the pipeline, build training datasets, run GRPO, create LoRA adapters, extract Claude sessions, convert Anthropic to OpenAI format, enrich OpenClaw tools, or work in the OpenClaw pipeline directory. |
Use this skill for the local fine-tuning pipeline under ~/openclaw/pipeline/.
Load references/datasets-inventory.md only when dataset variants, dataset
stats, or generated LoRA configs matter.
cd ~/openclaw/pipeline/
python extract_claude_sessions.py
python convert_anthropic_to_openai.py
python enrich_openclaw_tools.py
python build_training_dataset.py
Pipeline shape:
tool_use / tool_result into OpenAI tool_calls.mlx_lm_lora.train --config pipeline/training-data/openclaw-agentic/lora_config.yaml
Typical local assumptions:
5e-6cd ~/openclaw/pipeline/
python -m grpo.run --config grpo/config.yaml --dataset prompts.jsonl
Important files:
| File | Role |
|---|---|
grpo/config.py | GRPOConfig |
grpo/run.py | CLI entry point |
grpo/trainer.py | Training loop |
grpo/rollout.py | Group rollouts |
grpo/rewards.py | Reward functions |
grpo/loss.py | GRPO/PPO-style loss |
cd pipeline/
pytest test_extract.py -v
pytest test_convert.py -v
pytest test_enrich.py -v
pytest test_build.py -v
Proton CLI local email indexing and search pipeline.
Proton CLI local email indexing and search pipeline.
Expert LiteLLM - AI Gateway proxy multi-provider, routing fallback, budgets/spend, caching, guardrails, MCP gateway, logging. Consulter pour decisions d'integration LiteLLM dans IntelliSoins.
mlx-audio — Text-to-Speech (TTS), Speech-to-Text (STT) et Speech-to-Speech (S2S) optimisé pour Apple Silicon. Supporte le clonage de voix zero-shot (VoxCPM2, F5-TTS, Voxtral) et le fine-tuning (OuteTTS via mlx-lm, F5-TTS).
Manage the local AI/ML inference servers in this repository via `./aictl`. Use this project skill whenever Michael asks about ai-servers, aictl, local model server status, health checks, LaunchAgents, port conflicts, MLX servers, LiteLLM proxy, embeddings, reranker, GLiNER, Whisper STT, Kokoro TTS, oMLX, Ollama, or the PRO-G40 model-cache disk.
Stack AI/ML personnelle Michael Ahern, installée nativement sur le disque (Apple Silicon M3 Max via Homebrew + Launch...