Skill UI

Automates guidance on identifying and responding to model drift within ML deployment workflows, offering best practices, production-ready code, and validation support so you can keep inference systems stable.

nanoGPT is a simplified, minimalist implementation designed specifically for educational purposes. It reproduces the core architecture of GPT-2 (124M) using clean, hackable code, allowing users to understand the entire transformer pipeline from scratch. The framework supports full workflows, including data preparation on various datasets (e.g., Shakespeare, OpenWebText), model training, and text generation, making it ideal for students and researchers learning NLP and deep learning principles.

Guides building clinical AI integrations with the OpenEvidence SDK, covering singleton client setup, typed query builders, response transformers, and caching strategies for reliable decision support.

High-performance RLHF framework that combines Ray, vLLM, and ZeRO-3 to accelerate PPO/GRPO/RLOO/DPO training for 7B-70B+ models on distributed GPU clusters, providing a single-stack workflow for reward models and policy optimization.

Use this skill to implement OpenRouter function and tool calling when building agents or structured outputs with supported models such as GPT-4 or Claude, covering prerequisites, implementation guidance, testing, and production monitoring for API integrations.

Configure and operate OpenRouter's multi-provider support to compare models, select providers, and keep API-agnostic workflows running with consistent verification and monitoring steps.

Explain OpenRouter pricing, cost estimation, and efficient model selection to support budgeting and integration decisions. Use when estimating token expenses, aligning configurations, and validating API connectivity.

Outlines guarantees valid JSON/XML/code outputs by applying CFG/FSM-constrained token filtering, Pydantic schemas, and grammar-level sampling while supporting local Transformers, llama.cpp, and vLLM backends for fast inference.

Meta’s Prompt Guard is an 86M-parameter classifier that spots prompt injections and jailbreak attempts before they reach LLMs, filtering user input, third-party data, or RAG documents with multilingual support and sub-2ms latency for fast deployments.

Qdrant is a high-performance, Rust-powered vector database designed for production-grade RAG (Retrieval-Augmented Generation) and semantic search applications. It provides fast nearest neighbor search, supports complex hybrid search using metadata filtering, and offers horizontal scalability through sharding and replication. Ideal for building real-time recommendation systems and enterprise-level knowledge retrieval.

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.

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.