AI agent skill · by @aiming-lab

Nlp Pretraining

Best practices for language model pretraining and fine-tuning. Use when generating or reviewing NLP training code.

Archived 2026-10-23 · Source updated 2026-08-19 · License: MIT. Check the original repository for the latest version.

SKILL.md instructions

## NLP Pretraining/Fine-tuning Best Practice
Fine-tuning recipe:
- Use pre-trained checkpoints (HuggingFace hub)
- AdamW optimizer, lr=2e-5 to 5e-5
- Linear warmup (6% of total steps) + linear decay
- Batch size: 16-32 (use gradient accumulation for larger effective batch)
- 3-5 epochs for classification, 1-2 for generation
- Weight decay: 0.01

Parameter-efficient methods:
- LoRA: r=8-64, alpha=16-128, apply to q/v projections
- Prefix tuning: 10-20 prefix tokens
- Adapters: bottleneck dimension 64-256

Evaluation:
- Classification: accuracy, F1 (macro for imbalanced)
- Generation: perplexity, BLEU/ROUGE, human evaluation
- Use multiple seeds and report mean +/- std