Downloads · 30 days
0
Impliedhomeland/cosine-1B-60B-datamatched-resume
cosine-1B-60B-datamatched-resume is a machine learning model from Impliedhomeland. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Snapshot of the ONLY irreplaceable state from a node whose /tmp (ephemeral local SSD) is wiped by cluster maintenance. Everything else is deliberately NOT here because it is already published or is a documented rebuild.
Downloads · 30 days
0
Access
Public
Updated Aug 19, 2026
Repo size
850 GB
Likes
0
Public
Click a slice to open those files.
.pt874 GB · 100%
From the Hugging Face model README
Snapshot of the ONLY irreplaceable state from a node whose /tmp (ephemeral local SSD) is wiped
by cluster maintenance. Everything else is deliberately NOT here because it is already published
or is a documented rebuild.
ckpt/<budget>/<label>/<label>_60.00B_step30518.pt : the 60B pretrain final of each finished
code branch (weights only, final_model_only: true). ~15-20 h of 8xH100 each.posttrain/<budget>/<label>/posttrain/pycode/_sweep/lr<LR>/ : the Pycode SFT sweep, i.e.
result.json for all 12 LRs (the audit trail: val0/ep3/ep4, micro_batch, eval_micro_batch,
n_val_scored) plus ep4.pt for the best LR only (others pruned).| thing | where | cost |
|---|---|---|
| trunk + branch_A forks | Impliedhomeland/midtrain-bridge-1B-cosine-backbone | download |
| c4 / code / code_extra7B pools | Impliedhomeland/midtrain-bridge-data | download |
| 60.1B merged C4 pool | recipe in that dataset's pythia-70m/c4_60B/README.md | ~10 min |
| 14.0B merged code pool (budget18) | recipe in pythia-70m/code_combined14B/README.md | ~5 min |
| intermediate 5B-grid snapshots | not kept | nothing reads them |
latest.pt | not kept | the runs here are all complete |
handoffs/cluster_setup_h100.md in the repo).ckpt/ and posttrain/ under the same relative paths.cosine/datamatched/scratch/run_1B_60B_fill8.py. It is idempotent: branches with a
DONE marker are skipped, sweeps with result.json are skipped, and insert_row replaces
rather than duplicates. Completed rows are already in the repo's tables on /home (NFS).micro_batch 12 (world 8, H100 80GB); SFT train mb 8. Do not carry those to a different card.