Every official Hermes skill, in one place

Run PyTorch training across GPUs with minimal changes.

Embedding database for RAG and semantic search.

Zero-shot image classification and image-text search.

lm-eval-harness: benchmark LLMs (MMLU, GSM8K, etc.).

W&B: log ML experiments, sweeps, model registry, dashboards.

Fast vector similarity search at billion scale.

Speed up long-sequence transformer training and inference.

Constrain LLM output with grammars; guarantee valid JSON.

Fast BPE/WordPiece tokenization and custom vocab training.

llama.cpp local GGUF inference + HF Hub model discovery.

Outlines: structured JSON/regex/Pydantic LLM generation.

vLLM: high-throughput LLM serving, OpenAI API, quantization.

Structured LLM outputs validated with Pydantic.

On-demand GPU cloud instances for ML training.

Vision-language chat: VQA, captioning, image dialogue.

Serverless GPU cloud for ML jobs and model APIs.

HuggingFace hf CLI: search/download/upload models, datasets.

SAM: zero-shot image segmentation via points, boxes, masks.

Curate LLM training data: dedupe, filter, PII redaction.

OBLITERATUS: abliterate LLM refusals (diff-in-means).

Fine-tune large LLMs with LoRA on limited GPU memory.

Managed vector DB for production RAG and search.

Fully sharded data-parallel training for large models.

Clean training loops with built-in distributed support.

Vector search engine for production RAG systems.

DSPy: declarative LM programs, auto-optimize prompts, RAG.

Train sparse autoencoders to interpret model features.

Reference-free preference alignment, simpler than DPO.

RL post-training for LLMs with Megatron and SGLang.

Text-to-image generation, inpainting, and img2img.

High-throughput LLM inference on NVIDIA GPUs.

Pretrain LLMs at scale with PyTorch 4D parallelism.

Axolotl: YAML LLM fine-tuning (LoRA, DPO, GRPO).

TRL: SFT, DPO, GRPO, RLOO reward modeling for LLM RLHF.

Unsloth: 2-5x faster LoRA/QLoRA fine-tuning, less VRAM.

Transcribe and translate speech in 99 languages.