Grpo Python Packages

judgeval

The Continuous-Improvement Stack for Agents. Our environment data and evals power agent improvement and monitoring.

ms-swift

Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 600+ LLMs (Qwen3.8, DeepSeek-V4, GLM-5.1, InternLM3, Llama4, ...) and 300+ MLLMs (Qwen3-VL, Qwen3-Omni, InternVL3.5, Ovis2.5, GLM5.3, Gemma4, Llava, Phi4, ...) (AAAI 2025).

hud-python

RL environments + evals for AI agents. Define once, train anything.

hud

RL environments + evals for AI agents. Define once, train anything.

whileai

Scientific RL and SFT post-training for AI agents: build evals that can fail, simulate situations, judge checked against people, pass@1 with a 95% interval, train with GRPO/SFT/DPO, prove every gain on a held-out set. uv add whileai

zeroproof

Scientific RL and SFT post-training for AI agents: build evals that can fail, simulate situations, judge checked against people, pass@1 with a 95% interval, train with GRPO/SFT/DPO, prove every gain on a held-out set. uv add whileai

freesolo-flash-dev

LoRA post-training for open-weight models: SFT, GRPO, and on-policy distillation. Describe a run in TOML; Flash allocates a GPU, trains, streams checkpoints, and serves the adapter.

oumi

Easily fine-tune (SFT/RL), evaluate, and deploy Qwen, Gemma, or any open agentic LLM/VLM!

freesolo-flash

LoRA post-training for open-weight models: SFT, GRPO, and on-policy distillation. Describe a run in TOML; Flash allocates a GPU, trains, streams checkpoints, and serves the adapter.

trloom

YAML-driven end-to-end fine-tuning on Hugging Face TRL — one config for model, dataset, W&B, and Modal GPUs.

miniverl

verl for a single consumer GPU. PPO, GRPO and on-policy distillation on NVIDIA GPUs.

oat-llm

🌾 OAT: A research-friendly framework for LLM online alignment, including reinforcement learning, preference learning, etc.

forgelm

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

shadowlm

A fine-tuning SDK — any open model, any harness, any method. 12 training methods behind one argument; pure-stdlib core.

probe-the-harness

Setup checks for stale-data RL comparisons in language models

rldoctor

Diagnose why your GRPO/RLVR run is burning GPU-hours without learning. Reads the log you already have: W&B, TensorBoard, verl, TRL trainer_state.json.

rewardlint

Your RLVR reward function has a false-positive rate. Find out what it accepts that it shouldn't. Zero dependencies, no GPU, runs in a second.

areno

An easy-to-use, fast toolkit to scale up RL post-training on a single node.

textpolicy

Reinforcement learning for text generation on MLX (Apple Silicon)

gleamlm

GleamLM大语言模型全栈训练平台——核心代码用 PyTorch 从零实现，最低仅需 12GB 消费级显卡即可完整训练，另有一套工业代码与手写实现消费同一份数据逐层对照。模型训练支持CLI和可视化训练。

learnlens-rl

Universal evaluation layer for standard RL environments. Measures what an agent learned - not just how much reward it accumulated.

crucible-rl

Turn any real software into a replayable RL environment for training AI agents — deterministic replay, verifiable rewards, TRL & verifiers adapters.

oxrl

A lightweight post-training framework for LLMs and VLMs. 51 algorithms, 38 verified models. Scales with DeepSpeed, vLLM, and Ray.

zeroproof-simulations

Scientific RL and SFT post-training for AI agents: build evals that can fail, simulate situations, judge checked against people, pass@1 with a 95% interval, train with GRPO/SFT/DPO, prove every gain on a held-out set. uv add whileai