NLP Alignment Best Practices
Guides an agent through LLM alignment methods (RLHF, DPO, GRPO, SFT) and training pitfalls
- Repository
- aiming-lab/AutoResearchClaw
- Stars
- 14,802
- Licence
- MIT
- Files
- 1 file
What it does
This skill gives an AI agent a reference guide on how to align large language models, meaning training them to follow instructions and behave safely. Once installed, the agent knows the main approaches (RLHF with a reward model and PPO, simpler DPO, GRPO, and supervised fine-tuning), typical starting hyperparameters, a suggested training order, and common failure modes such as reward hacking, mode collapse, and catastrophic forgetting. It is written as static guidance in a Markdown file, not as a tool that performs training itself.
Great for
- ML engineers or researchers planning an LLM fine-tuning or alignment experiment
- Students learning the trade-offs between RLHF, DPO and SFT
- Agents helping with a research pipeline stage on model training or safety evaluation
Skip it if
- People who need a finished, tested training pipeline: the file only lists values and advice, and the hyperparameters are quoted without context, so they may not transfer to your model or data.
- Safety-critical decisions: the file names benchmarks such as TruthfulQA and BBQ but gives no evaluation procedure, so it cannot certify a model as safe.
Try asking
- “Explain the difference between RLHF and DPO and which one suits a small preference dataset.”
- “What learning rate and beta should I start with for DPO fine-tuning?”
- “List the common pitfalls to watch for when running PPO-based alignment.”
- “Outline a training recipe from SFT through preference optimization for my instruction-tuning project.”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 1 original files, the MIT licence and an attribution note.
nlp-alignment.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip nlp-alignment.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip nlp-alignment.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Treat the hyperparameters (DPO lr=5e-7, beta=0.1; PPO lr=1e-6, clip=0.2) as starting points and validate them on your own setup.
- Ask the agent to cite the referenced papers (Ouyang et al. 2022, Rafailov et al. 2023) when you need primary sources.
- Pair it with a concrete dataset or codebase so the agent can apply the general advice to your specific case.
What's inside
- SKILL.md
Redistributed unmodified from aiming-lab/AutoResearchClaw under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.