---

name: transformers description: Work with transformer models for CS research and development. Use when loading, fine-tuning, evaluating, or serving language, vision, audio, multimodal, embedding, or LLM-style models for tasks such as generation, classification, retrieval, instruction tuning, benchmarking, or agent backends.

Transformers

Use this skill for Hugging Face Transformer workflows in research, prototyping, and model-backed systems.

Common Uses

• run inference with pretrained checkpoints
• fine-tune encoders or generators on custom datasets
• benchmark LLMs or embedding models
• build retrieval, reranking, or classification pipelines
• configure tokenization, batching, and generation
• use open models as components inside agent or tool systems

Local References

Read only what matches the task:

• references/models.md
• references/tokenizers.md
• references/generation.md
• references/training.md
• references/pipelines.md

Working Rules

• Match model family to the actual task and latency budget.
• Make truncation, padding, and prompt formatting explicit.
• Record checkpoint version, tokenizer, and generation settings for reproducibility.
• Distinguish quick pipeline usage from training-grade code paths.
• If gated models or hub operations are involved, call out token requirements explicitly.

Default Output

When the user asks for transformer help without more detail, return:

1. model/task fit recommendation
2. inference or training setup notes
3. data and tokenization considerations
4. reproducibility or deployment caveats