OSP1-OLMo-3-7B-Think

Original archived OSP-1 checkpoints for Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning.

Code and recipes · Read-only paper snapshot · Retrained full OSP grid

Artifact identity

These three historical checkpoints (seeds 0, 1, 2; training price 2e-5) are the weights used by the paper's OLMo OSP-1 column. They were trained on the original labels. Natural-completion corrections were applied during evaluation; these weights were not retrained for that column. The retrained full-grid 2e-5 checkpoints are different artifacts and are hosted separately.

Each run includes original osp_best.pt, summary.json, available execution/migration metadata, and ARTIFACT.json. CHECKPOINTS.json records byte counts, original source paths, and SHA256; SHA256SUMS covers every supplied file except itself.

Loading

These custom top-two-block readers and stopping heads run alongside allenai/Olmo-3-7B-Think at revision d97e442d7cc678210054dbcc9b440894d62c89a4. They are not standalone language models. Download a run directory and pass it as --source-run with scripts/source_variants/olmo_osp_archived first on PYTHONPATH. See recipes.

The archived scorer does not install the trainer's fp32 boundary cast. Preserve the recorded runtime/device contract; its effect is not established by this publication. The known limitations also describe evaluation and data-availability qualifications. Base model weights, full question manifests, trajectory banks and policies are obtained separately.

Attribution

Source attribution is in NOTICE. This publication assigns no new project license; backbone and upstream terms continue to apply. Original checkpoint bytes are preserved.

Each run also includes its original val_scored.jsonl. The historical scorer requires this file and copies it into the scored output for calibration; row counts, byte counts and hashes are recorded under validation_scores in CHECKPOINTS.json. These records are the saved validation scores, not newly generated data.

Checkpoint inventory

Directory Seed Training price Role SHA256
archived-osp1/seed-0 0 2e-05 Archived OSP-1 c96df44639fa413d8ffebfe1593bb3ac7c726239403844e896830e796a11b04d
archived-osp1/seed-1 1 2e-05 Archived OSP-1 1d324ae77047baaebd7d027cc014b9aca1e59988fb021e49cd62bd9751ea2a5a
archived-osp1/seed-2 2 2e-05 Archived OSP-1 1c507173dbc5a383eb59cf4555369e24d18bd58beabef3f8560ad6728b3ad457
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for JingweiNi/OSP1-OLMo-3-7B-Think

Finetuned
(18)
this model