Skill UI

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.

This comprehensive workflow guides the end-to-end implementation of Retrieval-Augmented Generation (RAG) systems. It covers critical stages, including selecting optimal embedding models, setting up vector databases, designing chunking strategies, implementing hybrid retrieval, integrating Large Language Models (LLMs), and performing rigorous evaluation. Use this when building knowledge-grounded AI applications or advanced semantic search systems.

Automate Replicate AI workflows via the Composio MCP integration to inspect model schemas, run predictions (sync or async), upload input files, list versions, and manage prediction history and related assets.

Provides Meta AI’s Segment Anything Model for zero-shot image segmentation via interactive prompts or automatic mask generation, enabling rapid dataset creation and integration in vision pipelines.

Comprehensive skill set covering the full lifecycle of advanced visual AI systems. Expertise includes object detection (YOLO, Faster R-CNN), instance and semantic segmentation (Mask R-CNN, SAM), model training using PyTorch, and production deployment optimization using ONNX and TensorRT for edge or cloud targets.

A comprehensive, model-agnostic toolkit for advanced LLM development. It enables iterative prompt optimization through static analysis (token usage, clarity), measures critical RAG metrics (relevance, faithfulness, coverage), and validates complex agent workflows and configurations. Essential for rigorous LLM quality assurance and cost management.

Sentence Transformers provides 5000+ pre-trained models that generate sentence, text, and image embeddings for RAG, semantic search, clustering, and multilingual production workloads without relying on external APIs.

SimPO (Simple Preference Optimization) is a state-of-the-art, reference-free method designed for aligning Large Language Models (LLMs) using human preference data. It is an efficient alternative to DPO and PPO, notably outperforming DPO without requiring a separate reference model. It is ideal for practitioners who need faster, simpler, and more resource-efficient fine-tuning for preference alignment.

SAELens provides a framework for training and analyzing Sparse Autoencoders (SAEs). SAEs decompose the dense, often polysemantic activations of large language models into sparse, monosemantic features. Use this when you need to discover the discrete, interpretable concepts a model has learned, study feature superposition, or analyze specific safety-relevant behaviors (like bias or deception) within deep neural networks.

Speeds up LLM inference by combining speculative decoding with Medusa multiple heads and lookahead techniques so chatbots, code assistants, and other real-time workloads can run with 1.5-3.6× lower latency while preserving target-model quality.

A comprehensive guide to leveraging Stable Diffusion models via the HuggingFace Diffusers library. This skill facilitates state-of-the-art image creation, including Text-to-Image generation from prompts, Image-to-Image translation, inpainting, outpainting, and advanced workflow building. Supports various models like SDXL and SD 3.0, ideal for professional creative and computational vision tasks.

TensorBoard delivers a unified interface to track training metrics, debug network behavior, compare experiments, inspect model graphs, and profile performance for PyTorch or TensorFlow workflows.