Upload eval_push.py with huggingface_hub

auto ckpt ckpt_340k.pt

decontam-verify kept sample for Hart

Upload dump_kept_sample.py with huggingface_hub

Upload pod_prep.sh with huggingface_hub

remove mislabeled ckpt (was step-320000, autopush-race); canonical=ckpt_320k

remove mislabeled ckpt (was step-320000, autopush-race); canonical=ckpt_320k

remove mislabeled ckpt (was step-320000, autopush-race); canonical=ckpt_320k

remove mislabeled ckpt (was step-320000, autopush-race); canonical=ckpt_320k

Upload ckpt_autopush.py with huggingface_hub

Upload pod_prep.sh with huggingface_hub

Upload fabryka_push.py with huggingface_hub

Upload ckpt_autopush.py with huggingface_hub

Upload fork_launch.sh with huggingface_hub

Upload blend_edu.py with huggingface_hub

Upload build_decontam_index.py with huggingface_hub

Upload build_v2_blend.py with huggingface_hub

auto ckpt 120000

auto ckpt 80000

Upload v2b/ckpt_40k.pt with huggingface_hub

auto ckpt 80000

card: honest wiki correction (byte_ppl 2.372 / BPB 1.246 / eff ~75.9); maintainer re-benchmark PR#78; ARC-Easy 47.94 exact-match

auto ckpt 40000

card: dedup 64M load-recipe (single clean block)

glint_parity_eval: arch-aware loader (Qwen3 cfg from ckpt) - fix 64M maintainer load

card: add 64M Qwen3 load-recipe (maintainer audit load-safety)

Upload glint_parity_eval.py with huggingface_hub

auto ckpt 160000

card: 64M #6 flagship, board-protocol verified (glint bare-prompt: eff 77.51 / ARC 47.94)

revert card to pre-64M: hold 64M #5 pending board-protocol (glint bare-prompt) re-measure

card: add 64M #5 flagship (eff 77.70, board-native recompute-verified)

Upload eval/bpb_bin.py with huggingface_hub

Upload eval/board_eval_fabryka.py with huggingface_hub

auto ckpt 400000

auto ckpt 120000

auto ckpt 360000

auto ckpt 80000

auto ckpt 320000

auto ckpt 280000

auto ckpt 240000

auto ckpt 200000

auto ckpt 120000

v2-muon FWE/Cosmo ckpt 40k for board-native eval

forward-hold val vals/v2_mix_val.bin

forward-hold val vals/fwe_only_val.bin

arcmix-val cross-general forward-hold

model-def (GPT-ref Qwen3-arch) for same-harness loader

v1-muon arcmix ckpt 160k for same-harness re-measure

v1-muon arcmix ckpt 80k for same-harness re-measure

v1-muon arcmix ckpt 40k for same-harness re-measure