# Reinforcement learning fine-tuning

A static quantization version of UniReason-Qwen3-14B-RL, suitable for text generation and mathematical reasoning research scenarios.

ViGoRL is a vision-language model fine-tuned through reinforcement learning, used to clearly associate text reasoning steps with visual coordinates to achieve precise visual reasoning and positioning.

This is a fine-tuned version based on the Qwen-14B model, trained using the GRPO method, suitable for text generation tasks.

CodeV-R1-Qwen-7B is a model obtained through reinforcement learning fine-tuning based on the CodeV-R1 framework and the Qwen/Qwen2.5-Coder-7B-Instruct. It focuses on Verilog-related tasks and can effectively solve the problem of automatically generating hardware description languages in electronic design automation.

xGen-small is an enterprise-grade compact language model that achieves long-context performance with predictable low costs through domain-focused data curation, scalable pre-training, length extension, and reinforcement learning fine-tuning.

Phi-4-reasoning-plus is a large language model developed by Microsoft with enhanced reasoning capabilities, specifically optimized for complex mathematical problems and multi-step reasoning tasks.

OpenHands LM is an open-source coding model built on Qwen Coder 2.5 Instruct 32B, which performs excellently in software engineering tasks through special fine-tuning.

Language model fine-tuned using DPO method, suitable for text generation tasks

A Mist model optimized with Unsloth lazy-free framework and Huggingface TRL training library, achieving 2x training efficiency

Tülu 3 is a leading family of instruction-following models, offering fully open-source data, code, and training methodologies as a comprehensive guide to modern technology. Version 3.1 features improvements in the reinforcement learning phase, delivering enhanced overall performance.

A fine-tuned version based on the EleutherAI_pythia-1b-deduped model for generating concise summaries

An optimized model based on Meta LLAMA-3-8B, trained using lazy-free optimization techniques and the Huggingface TRL library, achieving 2x speed improvement

Official checkpoint of BLIP base model fine-tuned on MS-COCO dataset using MOCHA reinforcement learning framework to mitigate open-vocabulary description hallucination