Search Results for "recipes"

...In RLHF pipelines, reward models are responsible for evaluating generated responses and assigning scores that guide the model toward outputs that better match human preferences. The repository provides training recipes and implementations for building reward and preference models using modern machine learning frameworks. It supports multiple optimization strategies commonly used in alignment pipelines, including reinforcement learning with PPO, iterative supervised fine-tuning using rejection sampling, and direct preference optimization methods. ...

...The project focuses on the post-training stage of model development, where models are refined after pre-training to behave more helpfully, safely, and reliably in real-world applications. It provides detailed training recipes that explain how to perform tasks such as supervised fine-tuning, preference modeling, and reinforcement learning from human feedback. The handbook also includes reproducible workflows for training instruction-following models and evaluating alignment quality across different datasets and benchmarks. One of its goals is to bridge the gap between academic research on alignment methods and practical engineering implementation.

...The repo provides multiple ways to run models (CLI, web demo, and OpenAI-Vision–style APIs), along with quantization options that reduce VRAM needs (e.g., 4-bit). It includes checkpoints for chat, base, and grounding variants, plus recipes for model-parallel inference and LoRA fine-tuning. The documentation covers task prompts for general dialogue, visual grounding (box→caption, caption→box, caption+boxes), and GUI agent workflows that produce structured actions with bounding boxes.