slime — RL-дообучение LLM с Megatron-LM и SGLang

★ 7.7 · ai-tooling

slime is a Claude Code skill that guides LLM post-training with reinforcement learning using the slime framework from Tsinghua's THUDM team. The framework powers GLM-4.5, GLM-4.6, and GLM-4.7 by pairing Megatron-LM for actor training with SGLang for high-throughput rollout generation. It supports full parallelism (TP, PP, DP, SP), GRPO advantage estimation, asynchronous training with configurable buffer sizes, and multi-turn agentic workflows with tool use. Pre-built model scripts cover GLM-4.x, Qwen3, DeepSeek V3/R1, and Llama 3, with JSONL as the standard data format. The skill targets researchers and engineers who need scalable RL fine-tuning with native Megatron-LM integration and flexible data generation pipelines.

• #reinforcement-learning
• #post-training
• #grpo
• #megatron-lm
• #sglang
• #glm-models