
Orchestra-Research/AI-Research-SKILLs (MIT, v1.7.2, 9.9K stars) is a library of 98 SKILL.md files across 23 categories covering the full AI research lifecycle — fine-tuning, post-training, inference, evaluation, RAG, and more. The standout is the autoresearch orchestration skill: a dual-loop architecture that runs autonomous research end-to-end, routing to domain skills automatically.
references/ subdirectories that can run 300KB or more of deep documentation. 3 65 of the 98 skills carry "gold standard" status with comprehensive reference files. Total content across the library: ~130,000 lines.| Category | Skill count | Notable skills |
|---|---|---|
| Post-Training | 8 | TRL, GRPO, OpenRLHF, SimPO, verl, slime, miles, torchforge |
| Distributed Training | 6 | DeepSpeed, FSDP, Accelerate, Megatron-Core, Lightning, Ray Train |
| Optimization | 6 | Flash Attention, bitsandbytes, GPTQ, AWQ, HQQ, GGUF |
| Multimodal | 7 | CLIP, Whisper, LLaVA, BLIP-2, SAM, Stable Diffusion, AudioCraft |
| Fine-Tuning | 4–5 | Axolotl, LLaMA-Factory, PEFT, Unsloth |
| RAG | 5 | Chroma, FAISS, Pinecone, Qdrant, Sentence Transformers |
| Inference & Serving | 4 | vLLM, TensorRT-LLM, llama.cpp, SGLang |
| Prompt Engineering | 4 | DSPy, Instructor, Guidance, Outlines |
| Agents | 4 | LangChain, LlamaIndex, CrewAI, AutoGPT |
| Evaluation | 3 | lm-eval-harness, BigCode, NeMo Evaluator |
| Mechanistic Interpretability | 4 | TransformerLens, SAELens, pyvene, nnsight |
| Safety & Alignment | 4 | Constitutional AI, LlamaGuard, NeMo Guardrails, Prompt Guard |
| MLOps | 3 | W&B, MLflow, TensorBoard |
| ML Paper Writing | 2 | LaTeX workflow, Academic Plotting |
| Agent-Native Research Artifact (ARA) | 3 | ARA Compiler, Research Manager, Rigor Reviewer |

0-autoresearch-skill, which acts as a project manager for the other 97. Released in v1.4.0 (March 16, 2026), it implements a dual-loop architecture: 4research-state.yaml, research-log.md, findings.md, and organized subdirectories for literature/, experiments/, src/, data/, paper/, and a to_human/ queue. For literature search, it routes to Exa MCP, Semantic Scholar, arXiv, and CrossRef. 4/loop 20m command or OpenClaw's cron.add at a 20-minute interval — mandatory before any research work begins.npx @orchestra-research/ai-research-skillsnpx @orchestra-research/ai-research-skills install --allnpx @orchestra-research/ai-research-skills install post-training/plugin marketplace add orchestra-research/AI-research-SKILLs
/plugin install post-training@ai-research-skillshttps://www.orchestra-research.com/ai-research-skills/welcome.md with an instruction to follow the steps. 5 Note: the orchestra-research.com domain had a 0-byte response during collection, so the GitHub-hosted WELCOME.md is the reliable fallback.~/.orchestra/skills/ with symlinks into each detected agent's config directory--local flag: installs per-project with a .orchestra-skills.json lock file for team version controlnpm audit with 0 vulnerabilities. 6Run a full fine-tuning job on my dataset using Unsloth with 4-bit QLoRA.
Target: Llama-3-8B-Instruct. Dataset is in ./data/train.jsonl.
Minimize GPU memory — I'm on a single A100 40GB.03-fine-tuning category) provides Anthropic's Claude with complete setup commands, hyperparameter defaults, and common failure modes for Unsloth's 2× speedup path.Deploy my fine-tuned model for inference. I need to maximize throughput for batch API requests.
Model checkpoint is at ./checkpoints/llama-3-8b-ft. Expected QPS: 50.vllm serve command sequence with recommended --tensor-parallel-size and --gpu-memory-utilization settings for common hardware configurations.I want to investigate whether learning rate warmup schedules affect catastrophic forgetting
in continual fine-tuning of Llama models. Set up a research workspace and begin.
Use Exa for literature search. Run the inner loop for 10 experiments before synthesizing.findings.md and write a paper abstract from it. 4--all unless you have a clear reason to load the full library into every session.check-inventory.sh) that fails the build if documented counts diverge from actual SKILL.md files on disk — the current count of 98/23 is now CI-enforced. 6gh extension install — it hasn't been implemented. 15| Repo | Orchestra-Research/AI-Research-SKILLs |
| License | MIT |
| Version | v1.7.2 (June 16, 2026) |
| GitHub stars | ~9,900 |
| Supported agents | Claude Code, Cursor, Codex, Gemini CLI, Hermes, OpenCode, OpenClaw, Qoder, Qwen Code |
| Install | npx @orchestra-research/ai-research-skills install <category> |
| Maintainer | Orchestra Research (2-person team, Palo Alto) |
--all flag. 98 skills in context is overhead with no benefit unless you need coverage across the full AI research lifecycle in a single session.github.com
github.com
github.com
github.com
claudemarketplaces.com
explainx.ai
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.