npa-checkpoints

Noisy Policy Alignment — run checkpoints

Noise-policy (pi_psi) checkpoints from the noisy-policy-alignment project. The image generator g_theta is frozen and not included here; each file is the learned noise policy only.

Every checkpoint written during training: 220 files across 44 runs, 37.06 GB. Within a run, the highest step is the final policy; the lower steps are intermediate snapshots. Paths mirror noise_grpo/runs/<reward>/<run_dir>/, so the run key in the directory name encodes the hyperparameters (policy_type, K, G, B, lr, kl, reg_type, sigma, schedule, seed, steps).

Rows marked smoke are short debug runs, not experiments.