You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Download access is not open at this time.
Log in or Sign Up to review the conditions and access this model content.
Scaling-RL: Qwen3-4B RL checkpoints
Checkpoint backup. Qwen3-4B-Base trained with GRPO / PPO / DAPO, each under three rollout schedules (Sync / Tail / Sorted): 3 x 3 x 3 core runs plus single-seed variants.
Run matrix (4B)
- 27 core runs: 3 algorithms x 3 schedules x 3 seeds (
s1,s2,s3); 2K response cap, 8 responses per prompt, 128 prompts per update, 100 updates; checkpoints at updates 50 and 100. - 54 sweep runs: 3 algorithms x 3 schedules x 6 variants, one seed, 50 updates; checkpoint at update 50.
len1k/len4k(response cap 1K / 4K),n4/n16(responses per prompt),p64/p256(prompts per update). - Planned: same matrix and settings as 0.6B/1.7B. No checkpoints yet.
License
Not open for download at this time; access requests are not being approved.
Model tree for ZHEN-04/Scaling-RL-4B
Base model
Qwen/Qwen3-4B-Base