Skill UI

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.

This guide details a production-ready reference architecture for applications built on Groq's LPU inference API. It covers sophisticated model routing based on latency, quality, or cost needs, implementing streaming pipelines, and establishing robust multi-provider fallback chains (e.g., Groq primary to OpenAI backup). Ideal for designing and reviewing complex, resilient LLM systems.

Provides expert guidance for GRPO training with TRL, covering dataset prep, reward engineering, structured outputs, and multi-objective alignment when preference pairs are scarce.

A comprehensive reference architecture for implementing large-scale AI image generation using Ideogram. This system guides developers through building robust pipelines, including structured prompt templating (for brand consistency), utilizing multiple API endpoints (Generate, Edit, Remix), and implementing full asset workflow management from generation to final CDN delivery. Ideal for establishing brand asset systems and scalable AI integrations.

This skill serves as a comprehensive reference guide for Kling AI's various generative models. It details capabilities across video generation (T2V, I2V), image styling, lip sync, and virtual try-on. Users can compare model versions based on quality, speed, and cost to select the optimal model for their creative or production pipeline.

This comprehensive guide serves as a quick reference for diagnosing and fixing the most frequent errors and exceptions encountered when developing with LangChain. It covers issues such as import failures, API authentication problems, structured output parsing errors (Zod), agent looping, and version mismatches, providing root causes and copy-paste solutions for reliable LLM application development.

This guide outlines the comprehensive reference architecture for building advanced, AI-powered search and research applications, modeled after Perplexity. It covers three core tiers: a searchable service with model routing and caching, a citation pipeline for validation and rendering, and an orchestrator for multi-query research synthesis. Use this when designing complex, production-grade, search-augmented LLM integrations or reviewing advanced system patterns.

Provides implementation patterns and reference architecture for building advanced AI voice agents and automating complex phone call workflows using the Retell AI SDK. This skill is designed for developers requiring robust voice interaction, real-time telephony integration, and comprehensive conversational AI capabilities in their applications.

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.

This comprehensive guide outlines the production-grade reference architecture for advanced AI meeting systems using TwinMind. It details the end-to-end workflow, integrating high-accuracy transcription (Ear-3), AI summarization (GPT-4, Claude, Gemini), memory vault management, and automatic action item and calendar integration. Use this resource when implementing or managing a robust, automated AI meeting solution.

This guide details the production-grade reference architecture for running complex GPU compute workflows on Vast.ai. It covers the foundational three-tier pattern—Orchestrator, GPU Workers, and Artifact Storage—providing blueprints for job queue management, fault-tolerant training pipelines, and robust checkpointing. Use this when designing scalable and reliable ML/AI training systems.

Reference for building Graph Neural Networks with PyTorch Geometric, covering data objects, datasets, transforms, convolution layers, lazy initialization, and heterogeneous support so you can quickly tackle node, link, and graph tasks.