Skill UI

Evaluates large language models across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag) with standardized prompts and metrics, making it easy to benchmark HuggingFace or API models, compare releases, and track training progress.

Fine-tune LLMs with TRL’s RLHF toolkit, covering SFT, DPO, PPO/GRPO and reward model training so you can align HuggingFace Transformers with human preferences or feedback in post-training workflows.

Provides calibration-free weight quantization for LLMs down to 1-bit precision with multiple optimized backends, ideal for fast deployments with vLLM/HuggingFace and LoRA/PEFT fine-tuning without collecting calibration data.

The official Hugging Face Hub CLI (`hf`) is a comprehensive command-line tool for managing all types of resources on the Hugging Face Hub. Use it to download, upload, synchronize, and manage models, datasets, Spaces, and cloud buckets. It facilitates secure authentication, local cache management, and complex data operations, making it essential for ML workflows.

Execute complex and compute-intensive workloads directly on fully managed Hugging Face cloud infrastructure. This skill eliminates the need for local setup, allowing users to run jobs using cloud CPUs, GPUs, or TPUs. Ideal for large-scale data processing, batch inference, reproducible model training, and scheduled automation tasks, while persisting results to the Hugging Face Hub.

HuggingFace Accelerate adds distributed support to any PyTorch script with just four lines, providing a unified API for DDP, DeepSpeed, FSDP, and Megatron, plus automatic device placement, mixed precision, and simple launch/config commands for fast prototyping.

High-performance Rust-backed tokenization toolkit with Python bindings, offering BPE/WordPiece/Unigram, custom vocabulary training, alignment tracking, and production-ready integrations for NLP pipelines.

Provides expert guidance for using llama-factory to fine-tune LLMs through its WebUI, covering QLoRA/LoRA quantized setups, multimodal models, and automation tips for developers exploring models like Llama, Qwen, or Gemma.

LlamaGuard is Meta’s 7–8B safety-specialized LLM that filters both prompts and responses by classifying six threat categories, enabling fast inference via vLLM/SageMaker and integration into NeMo Guardrails for end-to-end moderation.

This skill provides comprehensive guidance on training Mixture of Experts (MoE) models, including architectures like Mixtral and DeepSeek-V3. It details techniques for achieving large-scale model capacity with significantly reduced computational costs (e.g., 5x reduction vs. dense models). Topics covered include top-k routing, load balancing, expert parallelism, and integration with frameworks like DeepSpeed and HuggingFace for efficient, resource-constrained training.

Covers HuggingFace PEFT workflows for LoRA/QLoRA-based LLM fine-tuning, enabling sub-1% parameter updates on limited GPUs plus adapter serving tips.

Quantizes HuggingFace LLMs to 8-bit or 4-bit with bitsandbytes, cutting memory by up to 75% while keeping accuracy, and supports QLoRA fine-tuning plus 8-bit optimizers for faster, memory-efficient training.