OSP-OLMo-3-7B-Think

Code and reproduction guide · Read-only paper snapshot

Original trained artifacts for Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning. Files retain their original bytes. CHECKPOINTS.json records the seed, training price where applicable, original source path, byte count and SHA256. Each run includes the original summary.json and available execution metadata. SHA256SUMS covers all supplied files except itself.

These are custom reader/head checkpoints used alongside the pinned base language model, loaded through the BELT repository's scorer. They are not standalone language models. No backbone weights, question manifests, full trajectory banks or fitted BELT stopping policies are bundled. See the code repository's recipes, asset inventory and known limitations.

Source attribution is in the repository's NOTICE. This artifact publication assigns no new project license; backbone and upstream terms continue to apply.

Backbone

allenai/Olmo-3-7B-Think at revision d97e442d7cc678210054dbcc9b440894d62c89a4.

Contents and use

The full OSP grid has five training prices (2e-6,5e-6,1e-5,2e-5,4e-5) and seeds 0, 1, 2. Each run contains osp_best.pt, the validation-selected top-two-block reader and stopping head. Use a run directory as --source-run with the matching iclr_score_pool snapshot; the base model is loaded separately.

This repository contains the 15 corrected-label full-grid checkpoints. The paper's OLMo OSP-1 column uses the 3 historical checkpoints in OSP1-OLMo-3-7B-Think, trained at 2e-5; do not substitute the retrained full-grid lambda-2e-5/ files.

Full-grid training began on GH200 and continued on Euler with recorded A100/RTX PRO 6000 devices and a path-only resume migration. Original run summaries and available migration records are included. Use scripts/source_variants/olmo_osp_full for the full-grid trainer/scorer and scripts/source_variants/olmo_osp_archived for the archived OSP-1 scorer. The archived scorer and trainer differ in fp32 boundary casting, and delivered code scoring includes different device paths; follow the recorded recipe and device receipts. Their effect is not established by checkpoint publication.

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
lambda-2e-6/seed-0	0	2e-06	OSP full-grid	b47661aa619c1fe9144ef9474ab5f4a110f99f1899d0acb14aa098b299548b58
lambda-2e-6/seed-1	1	2e-06	OSP full-grid	a1ccfcd11d70ccbb8dc3b068e0c3419f524c9f7962a6d6c6b5e7ae507c86f61f
lambda-2e-6/seed-2	2	2e-06	OSP full-grid	7b542bb965e92db16025cd9e4901b30d9b62ccdd099160edfa6c106e7364905a
lambda-5e-6/seed-0	0	5e-06	OSP full-grid	60cf73d04c1e0dd31207e4ee22fdb30c587e4cc15687264b18ccab1861862f82
lambda-5e-6/seed-1	1	5e-06	OSP full-grid	3324b31f432581b89bd9ff8196fe814c17d9143987a5e209f59730ebd9129e0c
lambda-5e-6/seed-2	2	5e-06	OSP full-grid	a060951b2f8ab9d0f112d22ecc9833ec73b7d64db75795696cf425976e028e3e
lambda-1e-5/seed-0	0	1e-05	OSP full-grid	b7674ced32429d53e9bda6d8fce6fe0eec99327e503208c52f0abc4692b6cf67
lambda-1e-5/seed-1	1	1e-05	OSP full-grid	e00fa28d2e8f66f2a557d5a8eb8189a2aac8e627c00a3a2bbd2141ec09eacece
lambda-1e-5/seed-2	2	1e-05	OSP full-grid	815cd191cb973fcfee455f6b4e4a1e5acdcdc4953a451ce8a126645020c4967a
lambda-2e-5/seed-0	0	2e-05	OSP full-grid	3801940325cc451c834cdccb5da2460607c8f40514ae6b52e6293e1662e9468a
lambda-2e-5/seed-1	1	2e-05	OSP full-grid	06293b61381feaf47440b9ea8dad1af1d8f622f1f53b65a68b58d30d1b97f3a5
lambda-2e-5/seed-2	2	2e-05	OSP full-grid	681fc94ba90032085ad6f020ab00ccf4a48c97927ae1b8962982d9647775e5c2
lambda-4e-5/seed-0	0	4e-05	OSP full-grid	62d95e96ab031b563655e95a55f880519ee908a94337cbe5b3daa51806acf228
lambda-4e-5/seed-1	1	4e-05	OSP full-grid	a7716dfb7cb3a9a3306233967c20cc54a822e560560cd4e9813955da8dcac4c4
lambda-4e-5/seed-2	2	4e-05	OSP full-grid	e3434a04d462dd81815564152841c09b1f4d42d075538ae712b04c0487080966