jingwenliu-HF/self-improve-len-gen

Model zoo

Handoff model checkpoints + eval stats, organized by puzzle domain:

subtree	domain	status
rubik2x2/	2×2×2 Rubik's cube	populated (5 of 6 settings; the 6th, optimal_data_near_solved_goal, is still training)
puzzle15/	15-puzzle	populated (all 6 settings; base + RFT rounds d16–d25, 5 seeds)

Each subtree groups models by training setting, with a pretrained base/ and iterative RFT rounds per seed. See each subtree's README.md for the full layout and per-setting coverage.

Checkpoints

All checkpoints (.pt) are included directly in this repo alongside their eval stats — there is nothing extra to download. Each setting folder holds a pretrained base/ plus iterative RFT rounds, per seed. See each subtree's README.md for the full layout and per-setting coverage.