hkust-nlp/simpleRL-reason

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

3874 stars · 285 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

A research codebase from HKUST-NLP implementing a simple reinforcement learning recipe (rule-based rewards on GSM8K/Math data) to train LLMs' reasoning abilities from base models without SFT. It includes training/eval code, model checkpoints for 10 base models, and analysis tools accompanying the SimpleRL-Zoo paper.

Use cases

• train an LLM with reinforcement learning to improve math reasoning
• run zero RL training on base models like Qwen2.5 or Llama3
• reproduce SimpleRL-Zoo paper results
• compare reasoning behaviors across models during RL training
• fine-tune a 7B model with only 8K examples and rule-based rewards
• analyze response length and accuracy trends during RL training

When to choose

• you want a minimal, proven RL recipe for eliciting reasoning in base LLMs
• you need to reproduce or extend the SimpleRL-Zoo results
• you want to train with rule-based rewards instead of a reward model
• you're researching emergent reasoning behaviors during RL

When to avoid

• you need production-grade, supported LLM training infrastructure
• you want RLHF with human preference data or reward models
• you need SFT pipelines or general instruction tuning
• you lack multi-GPU resources for large model training

Facets

library · maturity active

llm-training reinforcement-learning machine-learning large-language-models machine-learning deep-learning python rlhf rule-based-reward zero-rl math-reasoning verl model-training research-code natural-language-processing gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="hkust-nlp/simpleRL-reason")