olmo-3-7b-jlens (placeholder — fits in progress)

This repo will host fitted Jacobian lenses (jlens) for OLMo-3-7B checkpoints not covered by Neuronpedia's published set (which currently has olmo-3-1025-7b, the base checkpoint, but no post-training stages).

Status: fitting in progress. Nothing is uploaded yet. Lenses will land here once fitted and validated against a known-answer test (see jlens-toolkit) confirming they agree with Anthropic/Neuronpedia's own fitting methodology at the layers where the fit has converged.

Planned contents

• Olmo-3-7B-Instruct_jacobian_lens.pt — post-RL checkpoint
• Olmo-3-7B-Instruct-SFT_jacobian_lens.pt — SFT-stage checkpoint
• Convergence CSVs + fit config, matching Neuronpedia's own release format
• A known-answer validation report (per-layer agreement with the published base-checkpoint lens's fitting methodology)

A note on reliability by layer

Jacobian-lens fits are Monte-Carlo averages, not trained/optimized objects — but convergence is layer-dependent, and shallow layers can remain under-converged even at large prompt counts (n=1000) while deep layers converge tightly. Any release here will state which layers are validated vs which should be treated as provisional, rather than a single blanket reliability claim. See the toolkit repo's README for the known-answer test that surfaces this.

Credits

Anthropic (jlens method/code, Apache-2.0), Neuronpedia (published lens catalog + fitting conventions), Allen Institute for AI (OLMo checkpoints).