Orchestra's 98 Research Skills Give Your Agent the Full ML Lifecycle

ML research means juggling dozens of specialized tools, and the Orchestra Research README's diagnosis is blunt: researchers spend more time debugging infrastructure than testing hypotheses. Its answer is this 13k-star library — the README calls it the most comprehensive open-source skills library enabling AI agents to autonomously conduct AI research, from idea to paper.

Why This Skill Matters

The inventory, reconciled repo-wide in v1.7.1 with a CI drift guard that fails when documented counts diverge from disk, stands at 98 skills across 23 categories. The coverage maps to a real research loop: 4 fine-tuning skills (Axolotl, LLaMA-Factory, PEFT, Unsloth), 8 post-training skills (TRL, GRPO, OpenRLHF, SimPO, verl, slime, miles, torchforge), 6 distributed-training skills, 4 inference skills (vLLM, TensorRT-LLM, llama.cpp, SGLang), 5 RAG skills, 7 multimodal skills, plus evaluation, mechanistic interpretability, MLOps, ML paper writing, and ideation.

Sitting on top is autoresearch, a single orchestration skill using a two-loop architecture — an inner optimization loop and an outer synthesis loop. It manages the lifecycle from literature survey to paper writing and routes to domain skills automatically, so the agent does not need to know which domain skill to invoke. The demos gallery includes two papers the README says were produced end-to-end by agents using it: one that discovered norm heterogeneity predicts fine-tuning difficulty, and one that analyzed DPO as a rank-1 perturbation.

Installation

One command runs an interactive installer that auto-detects your installed coding agents (the README lists Claude Code, Hermes Agent, OpenCode, Qoder, Cursor, Gemini CLI, among others):

npx @orchestra-research/ai-research-skills

Skills install to ~/.orchestra/skills/ with symlinks into each agent (falling back to copies on Windows), and you can take everything, a quickstart bundle, a category, or individual skills. Two management commands cover upkeep:

npx @orchestra-research/ai-research-skills list
npx @orchestra-research/ai-research-skills update

Or hand the whole library to your agent with one instruction:

Read https://www.orchestra-research.com/ai-research-skills/welcome.md and follow the instructions to install and use AI Research Skills.

Real Workflow: Install One Category via the Claude Code Marketplace

Full libraries are a lot of context. If you only need fine-tuning, the marketplace route installs a single category:

/plugin marketplace add orchestra-research/AI-research-SKILLs
/plugin install fine-tuning@ai-research-skills

Other categories follow the same pattern — post-training@ai-research-skills, inference-serving@ai-research-skills, distributed-training@ai-research-skills.

Real Workflow: Ask for a Fine-Tuning Run

With the skills loaded, pose the task the way the README's use cases do:

I need to fine-tune Llama 3 with my custom dataset. Use the Axolotl skill
to set up a YAML config, then walk me through launching training.

Each skill follows the same structure: a 50–150 line SKILL.md quick reference, backed by a references/ directory with deep documentation the README says is sourced from official repos, real GitHub issues, and production workflows.

Tips

• Install by category rather than all 98 unless you are running autonomous research — the autoresearch layer is what consumes the full inventory.
• The README's own stats table reports an average of about 420 lines per skill, so even individual skills carry more depth than a cheatsheet.
• Every skill is synced to the Orchestra Research site for one-click adds, if you prefer browsing over installing.
• Individual skills may reference libraries with their own licenses — the README asks you to check each project's license before use.

When Not to Use This

This is a research and ML-engineering library, not general web development tooling — if your task is shipping an app, the skills here are beside the point. Framework knowledge also moves fast: the skills document tools as of their last sync, so for a brand-new library release the official docs may be ahead. And if you want a curated demo of one deep tool rather than broad lifecycle coverage, a single vendor's docs will be a tighter read.

See the leaderboard for more skills.