Dataset of hackable TerminalBench-style tasks and exploit trajectories

[ICML 2026] Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks

🔥 LLM-powered GPU kernel synthesis: Train models to convert PyTorch ops into optimized Triton kernels via SFT+RL. Multi-turn compilation feedback, cross-platform NVIDIA/AMD, Kernelbook + KernelBench

slime is an LLM post-training framework for RL Scaling.

A version of verl to support diverse tool use [TMLR 2026]

[COLM 2025] Official repository for R2E-Gym: Procedural Environment Generation and Hybrid Verifiers for Scaling Open-Weights SWE Agents

[EMNLP 2025] LightThinker: Thinking Step-by-Step Compression

[ICLR 2025] DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference

[ICLR 2025] DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference

ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning & ReCall: Learning to Reason with Tool Call for LLMs via Reinforcement Learning

[ACL2026] "MiniRAG: Making RAG Simpler with Small and Open-Sourced Language Models"

Reproducing R1 for Code with Reliable Rewards

Based on the R1-Zero method, using rule-based rewards and GRPO on the Code Contests dataset.

Docker image NVIDIA GH200 machines - optimized for vllm serving and hf trainer finetuning

RL Scaling and Test-Time Scaling (ICML'25)

Moatless Testbeds allows you to create isolated testbed environments in a Kubernetes cluster where you can apply code changes through git patches and run tests or SWE-Bench evaluations.

Minimal reproduction of DeepSeek R1-Zero

A minimal language for Isabelle/HOL, designed for easing machine learning.

🤗 smolagents: a barebones library for agents that think in code.

珠算代码大模型（Abacus Code LLM）

Code to compute AnthroScore, a computational linguistic measure of anthropomorphism in text

AllenAI's post-training codebase

Super-Efficient RLHF Training of LLMs with Parameter Reallocation

Formatron empowers everyone to control the format of language models' output with minimal overhead.