examples

Examples

End-to-end usage examples for prime-rl, referenced from the top-level README.

• basic/ — walk-throughs for the core environments on 1–8 GPUs (baseline eval → optional SFT warmup → RL → eval), each with its own README:
  • reverse-text — smallest end-to-end loop (single-turn, 0.6B)
  • alphabet-sort — multi-turn, user simulator, LoRA
  • wiki-search — multi-turn tool calling, LoRA
  • wordle — multi-turn (~6-turn games)
  • hendrycks-sanity — single-turn math, long-running
• advanced/ — larger, mostly multi-node runs on frontier models, one folder per model: qwen3-30b-a3b (math/swe/tool), glm-4.5-air (search/swe/terminal), glm-5.2 (large-scale + PD-disaggregated inference), minimax-m2.5 (swe), nemotron-3-super (swe), intellect-3.1 (swe).

Dev-sized (2-GPU) counterparts of basic/ live in configs/basic/.