Skill UI

Compress large language models quickly using Wanda, SparseGPT, and N:M pruning to enforce hardware-friendly sparsity, enabling faster inference and deployment without retraining.

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.

A comprehensive guide for training or fine-tuning advanced sentence-transformer models. This skill assists in selecting the correct architecture—including Bi-encoders (SentenceTransformer), CrossEncoders (for reranking), and SparseEncoders—for various Natural Language Processing (NLP) tasks. Use cases include semantic retrieval, similarity matching, document clustering, and pair-wise classification.

Train or fine-tune SentenceTransformer, CrossEncoder, and SparseEncoder models for retrieval, similarity, clustering, classification, reranking, and embedding tasks. This skill provides production-ready templates with proper loss functions, evaluators, and training configurations for various NLP tasks including bi-encoder, reranker, and sparse vector models.

Learn to train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. This skill covers sparse architectures like Mixtral 8x7B, reducing compute costs by 5x compared to dense models. It includes routing mechanisms, load balancing, expert parallelism, and inference optimization for scaling large-scale AI models efficiently.