data/README.md

This directory contains data of the following:

1. SmolLM + GSM8k

• Model: SmolLM_MaxRL_rollout128_step1000
• Dataset: 512 prompts x 512 rollouts in GSM8k Training set

2. Qwen Math

• Model: Qwen3-1.7B_MaxRL_1000steps
• Dataset: 512 prompts x 512 rollouts in Polaris-53K dataset
• Model: Qwen3-4B_MaxRL_1000steps
• Dataset: 1024 prompts x 128 rollouts in Polaris-53K dataset

3. Maze

• Models: Maze17_bz256_ns64(0,5000,8000steps)
• Dataset: Maze17_train_parquet, 2048 prompts x 1024 rollouts for each prompt