grpo-training

official github code for "SmartPhotoCrafter: Unified Reasoning, Generation and Optimization for Automatic Photographic Image Editing"

[ACM MM 2026] ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination

Group-relative Trajectory-based Policy Optimization: Increasing Quality and Training Stability

OpenMedRL is an open-source reinforcement learning environment for benchmarking LLM-powered medical agents in emergency care. It simulates triage, dynamic patient progression, resource constraints, and uncertainty-aware clinical decision-making.

RL training environments with verifiable rewards for coding agents. Works with TRL, Unsloth, verl, OpenRLHF.

An OpenEnv RL environment where an LLM agent plays the buyer and negotiates against an LLM-powered seller over real marketplace listings.

This repository contains my personal notes and hands-on implementations for fine-tuning and post-training Large Language Models (LLMs).

GRPO training that runs until you stop it on a single RTX4090 with vllm 0.26.0 (Linux Only).

GRPO-based RL fine-tuning of Qwen3.5-4B for reliable function calling, inspired by FunRL. Trained on a single free Kaggle T4 GPU via Unsloth 4-bit QLoRA. Boosts unseen BFCL parallel tool-calling accuracy from 0.33 to 0.73.

A framework for training/evaluation FastWAM with SFT/RL and simulation environment.

Low-cost GRPO fine-tuning pipeline for valence-arousal state estimation and tone-aligned responses in Llama 3.2 1B.

A reinforcement learning fine-tuned model that generates Linux terminal commands from natural language descriptions. Trained using GRPO (Group Relative Policy Optimization) on a custom terminal task environment inspired by CAMEL-AI's SETA framework.

[SIGGRAPH Asia (TOG) 2026] Official Implementation of "ABACUS"

Train a small reasoning model to obey an explicit thinking-token budget ("Think for maximum N tokens") using GRPO. Reproduces the L1/LCPO recipe (Aggarwal & Welleck, CMU 2025) on DeepSeek-R1-Distill-Qwen-1.5B, then extends it with a GGUF release and a llama.cpp demo of the budget knob.