MoS-Aurora-Experiment-Archive

ryan-0608/MoS-Aurora-Experiment-Archive

MoS-Aurora-Experiment-Archive is a machine learning model from ryan-0608. Use it for the machine learning task on the model card, and read the license before you ship it in a product.

Latest supplemental cleanup: file-level dependency review and receipts. Total local inode-exclusive cleanup is now 365.66 GiB; current experiment assets remain protected.

Downloads · 30 days

Access

Public

Updated Oct 20, 2026

Repo size

2.4 TB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors1.2 TB · 72%

At a glance

Access: Public
Created: Aug 6, 2026
Updated: Oct 20, 2026
SHA: c7085523

Model card

From the Hugging Face model README

MoS Aurora Experiment Archive

Latest supplemental cleanup: file-level dependency review and receipts. Total local inode-exclusive cleanup is now 365.66 GiB; current experiment assets remain protected.

Public research archive for Aurora, DFlash and MoS experiments: selected model-only checkpoints, serving exports, configurations, logs, evaluation records and historical evidence. An archive entry does not imply a positive result; failed, cancelled and retracted results retain their original status.

Archive index

2026-09-15 verified cleanup catalog · 2026-09-08 complete closeout catalog

The directory counts below are a payload snapshot at revision 1eaebd9850e19799cba7a931848f8071125ebbed (2026-09-14T22:52:24.150884+00:00). They describe logical archived files, not local disk usage or an assurance that every current experiment is backed up.

Directory	Files	Logical GiB
100k_token_v1_vs_dense_8pass	243	12.009
ablation_4b_800k_20260825	673	50.108
ablation_4b_800k_round4_20260827	25	14.821
blockpos_4b_self800k_20260908	24	11.354
cancelled_code_saturation_4b_20260909	132	13.860
code_overfit_4b_20260909	166	0.080
code_overfit_4b_dense_20260909	410	27.369
code_overfit_4b_shared_code_20260909	461	41.499
d0_online_token_mos_dense_4pass_20260815	120	50.174
data_clean_4b_trajectory_v2_cancelled_20260912	18	0.047
dflash_27b_mos_experts	15	16.114
e5_4b_800k_20260831	345	115.884
e5_8b_800k_20260831	407	281.347
e5self_4b_800k_20260905	2,311	526.309
e8_4b_800k_20260825	88	11.880
e8_8b_800k_20260827	204	106.919
e8_interleave_8b_100k_20260823	115	21.561
expert_probe_20260818	32	0.200
generalist_init_m3_8pass	36	17.580
online_b1_sequential_warm	18	3.404
router_granularity_p1	155	65.325
router_granularity_p1_online	617	86.785
router_morphology_4b_20260911	55	0.336
router_rank_e1_20260913	123	41.629
router_rescue_4b_k4_e1_20260912	58	8.420
specforge_8b_800k_clean_20260827	97	17.149
storage_cleanup_20260915	568	27.311
storage_closeout_20260908	7,809	1259.383

Related archives

Restore and interpretation

Use the per-experiment config, protocol, exact epoch/step and stored revision. Model-only checkpoints and HF exports do not restore optimizer, learning-rate scheduler or RNG state. Several historical experiments include implementation smokes or intermediate checkpoints; use their own README and result evidence to distinguish them from scientific endpoints.

Local cleanup status is separate from Hub retention. Existing historical Hub prefixes were preserved during the 2026-09-15 index reconciliation; no old remote model directories were deleted or renamed.

Created: Aug 6, 2026
Updated: Oct 20, 2026