Guide · Agent skill
How to use the NLP Pretraining Best Practices agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The nlp-pretraining skill is a set of written guidelines that an AI agent reads when you ask it to write or review code for training or fine-tuning language models. It gives the agent standard recommended settings for learning rate, batch size, warmup, weight decay, LoRA and evaluation. This guide explains what the skill changes, how to install it, how to trigger it with plain requests, and where its limits are.
NLP Pretraining Best Practicesaiming-lab/AutoResearchClaw · MIT Download the skill Quick start
- 1Download the ZIP. Download the nlp-pretraining skill ZIP from this page.
- 2Unzip the skill. Unzip the file so you have a folder containing SKILL.md.
- 3Place the folder. Move the folder into ~/.claude/skills/ for personal use, or into .claude/skills/ in your project.
- 4Or upload in the apps. Alternatively, upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- 5Make a request. Ask the agent to write or review a fine-tuning or pretraining script, naming your model and task.
- 6Check the output. Compare the chosen settings against the recipe and confirm results are reported across multiple seeds.
What does the nlp-pretraining skill make the agent do?
The skill contains no executable code. It is a SKILL.md file of guidelines. When it is active, the agent uses these guidelines as defaults while it writes or reviews training code, so its advice and code tend to follow common practice.
The recipe covers fine-tuning defaults such as pre-trained checkpoints from the HuggingFace hub, the AdamW optimizer with a learning rate of 2e-5 to 5e-5, linear warmup over 6% of total steps followed by linear decay, and weight decay of 0.01. It also covers parameter-efficient methods (LoRA, prefix tuning and adapters) and evaluation advice such as using multiple seeds and reporting mean +/- standard deviation.
- Fine-tuning: AdamW, learning rate 2e-5 to 5e-5, batch size 16-32, 3-5 epochs for classification and 1-2 for generation.
- Parameter-efficient: LoRA with r=8-64 and alpha=16-128 on q/v projections; prefix tuning with 10-20 prefix tokens; adapters with bottleneck dimension 64-256.
- Evaluation: accuracy and macro F1 for classification (macro for imbalanced data); perplexity, BLEU/ROUGE and human evaluation for generation.
How do you install the nlp-pretraining skill?
Download the skill ZIP from this page and unzip it. For a personal install, place the unzipped skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder instead.
You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
How do you trigger the skill with a natural request?
You do not need a special command. The skill is meant to be used when you ask for NLP training code to be generated or reviewed. Mention the kind of model and the task, and words such as fine-tuning, BERT, GPT, LoRA or pretraining help match the request to the skill.
Be specific about what you want reviewed or written. The skill gives defaults, so the more context you give about your model and data, the more useful the adjusted advice will be.
- Write a fine-tuning script for BERT on a text classification dataset using the usual defaults.
- Review this training script for a GPT-style model and tell me if the learning rate, warmup and weight decay look reasonable.
- Add LoRA to my transformer fine-tuning code and suggest rank and alpha values.
- Check whether my evaluation for this classifier follows common practice, including seeds and macro F1.
What examples show the skill in use?
Reviewing an existing setup is a good first test. Paste your training configuration and ask the agent to compare it against the recommended ranges. You can then decide which differences matter for your experiment.
For new code, ask for a script and check the defaults it chose against the recipe. Confirm that it reports results across several seeds rather than a single run.
What does the skill need, and what are the safety notes?
The skill lists no requirements. It does not run scripts and does not use the network. It only shapes what the agent writes and says.
The skill does not install or run anything, and it gives generic ranges rather than tuned values for your dataset. Any code the agent writes still needs your review and testing before you rely on it. The agent may still produce wrong code.
- Requires: nothing listed
- Runs scripts: no
- Uses network: no
- Do not use the output alone for production decisions that need validated results.
What tips help you get better results?
Treat the recommended ranges as starting points. Tune learning rate, batch size and epochs on a validation set for your own data.
Use gradient accumulation if you want a larger effective batch size than your hardware allows, as the recipe suggests.
- Give the agent your model name, task and dataset size.
- Ask for results averaged over multiple seeds with mean +/- std.
- Check the generated code runs on a small sample before a full training run.
What people use it for
- Drafting a BERT fine-tuning script for text classification with standard learning rate and warmup.
- Reviewing a GPT-style training config for weight decay, batch size and epoch count.
- Adding LoRA to an existing transformer script with suggested rank and alpha ranges.
- Checking that a classifier evaluation reports macro F1 and mean +/- std across seeds.
Frequently asked questions
How do I install the nlp-pretraining agent skill?
Download the skill ZIP from this page and unzip it. Put the folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
Is the nlp-pretraining skill free to use?
The item is listed under the MIT licence. Check the licence file in the source repository for the full terms before reuse.
What does the nlp-pretraining skill require?
The skill lists no requirements. It does not run scripts and does not use the network, so it adds guidance rather than installing software.
Can the skill train a model for me?
No. It contains only written guidelines. It does not install or run anything, so you still need a working training pipeline.
Are the hyperparameters in the skill guaranteed to work?
No. The recipe gives generic ranges, such as a learning rate of 2e-5 to 5e-5, not values tuned for a specific dataset. Validate them on your own data.
How is this different from just asking an agent for training code?
Without the skill the agent uses its general knowledge. With it, the agent follows a fixed set of recommended defaults for optimizers, warmup, LoRA and evaluation, so its advice is more consistent.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are writing or reviewing fine-tuning code for BERT- or GPT-style models and want the agent to start from common defaults and evaluation practice. Pick something else if you need validated results for production, or if you want a finished model without working with training code, since this skill only shapes advice and does not run anything.
Download the skill