VIP Qwen3-4B math — MSE value loss, AP + active sampling
Checkpoint branches from Hyak Slurm job 37354948. Each completed 100-step checkpoint is sanitized and published as revision step-N.
Training uses 32 accepted prompts × 8 rollouts, active sampling, zero-variance filtering, answer-prefix value conditioning, MSE value loss, decoupled GAE, and SAE.
Model tree for hamishivi/vip-g8-p32-mse-ap-as-zvf-qwen3-4b-math
Base model
Qwen/Qwen3-4B-Base