What problem does it solve?
This Skill automates the post-training process for large language models (LLMs) using the slime framework, providing a streamlined approach to reinforcement learning (RL) and scaling Megatron-LM for RL applications.
Core Features & Use Cases
- LLM Post-Training: Facilitates the use of the slime framework for post-training of LLMs, enabling RL enhancements.
- Megatron-LM Integration: Offers seamless integration with Megatron-LM for RL scaling.
- Custom Data Generation: Supports custom data generation workflows for flexible prompt management and sample storage.
- Use Case: Ideal for training GLM models, implementing custom data generation, or needing tight Megatron-LM integration for RL scaling.
Quick Start
Run the following command to start a slime training session:
python train.py --actor-num-nodes 1 --actor-num-gpus-per-node 4 --rollout-num-gpus 4 --advantage-estimator grpo --use-kl-loss --kl-loss-coef 0.001 --rollout-batch-size 32 --n-samples-per-prompt 8 --global-batch-size 256 --num-rollout 3000 --prompt-data /path/to/data.jsonl