Automate end-to-end RL training workflows for language models with SLIME.
npx skills add https://github.com/yzlnew/infra-skills --skill slime-user
Please help me install this Agent Skill. Skill: slime-user Source: https://github.com/yzlnew/infra-skills/tree/main/slime-user Command: npx skills add https://github.com/yzlnew/infra-skills --skill slime-user
SLIME User Guide provides a structured, extensible workflow for RL training of language models, unifying rollout, reward modeling, and backend integrations to reduce setup complexity and accelerate experimentation.
Launch a SLIME training session with a minimal setup and provide a dataset and HF checkpoint path.
High-intent search queries and answers about installing and using this skill.
Multi-turn RL training with tool calling is orchestrated by automating rollout, reward modeling, and backend integrations. This workflow supports custom generation and reward functions for interactive language model training.
Distributed RL training using Megatron or FSDP backends is set up through a unified argument schema. This configuration supports scalable training across single or multi-node clusters while loading HF-backed model checkpoints.
Yes, dynamic sampling and partial rollout can be used to improve sample efficiency across datasets during RL training. This approach optimizes data management with JSONL prompts, labels, and metadata for RL tasks.
Yes, SLIME supports modular customization of generation and reward functions alongside troubleshooting hooks for bespoke RL workflows. This extensibility reduces setup complexity and accelerates experimentation.
JSONL data formatting is required for prompts, labels, and metadata to launch a minimal RL training session. You must provide a dataset and an HF checkpoint path to initiate the SLIME training workflow.
thinking-machines-labConfigure and run Direct Preference Optimization training with a model and preference dataset.
attentiondotnetImplement GRPO/TRL reinforcement learning fine-tuning workflows for language models.
KarlinskySAdd distributed training to PyTorch scripts with mixed precision and multiple parallelism backends.
SignmanalCoordinate distributed Megatron-LM training and SGLang inference for RL post-training.
