Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('validation'): ('text', {}), NamedSplit('test'): (None, {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

ALG-HODGE-27B-01: Bayesian evidence geometry in Qwen3.5-27B

Public archive of experiment exp_01m6k3brj2fhx8y4b9hgd57pxc, using Qwen/Qwen3.5-27B revision fc05daec18b0a78c049392ed2e771dde82bdf654, not the Qwen3.8-27B model used in the separate M-VOI-01 dataset. All 693 retained durable objects are intended to be represented at their original relative paths under alg_hodge_27b_01/ and review_contract/; authored research files are under research_source/. Read the separate publication manifest and verification receipt for completion and SHA-256 evidence.

Authoritative repaired experiment and review

alg_hodge_27b_01/repaired_v2/ is the authoritative repaired result, with 512 family-disjoint Bayesian tasks (256 train, 128 calibration, 128 evaluation), 40,448 contexts, 128 captured activation shards, six-layer Goodfire stores, frozen readout weights selected_weights.npz (not model checkpoint weights), full-precision readouts and intervention rows. alg_hodge_27b_01/review_v3/ is the final review; review/ records earlier integrity and example audits, review_v2/ paired-control corrections, and review_contract/plot_ready_worker_v3.json the last verified plot-data correction. These review versions are retained in place to preserve the audit trail. The evaluation cases and all results are exposed here, not an untouched holdout for future confirmatory work.

The report is research_source/worktree/REPORT.md (also in the researcher's source folder); source, tests, run manifest and figures are under research_source/. The accepted outer experiment source snapshot was commit 7d6cb23b4e00a9fdb7f6685673da8c3759a0d2d2. The researcher repository's research/alg-hodge-27b-01 branch was staged but never committed or published: its base was fab953c987b3b39a2a3780400adb608daaaa3a7d; files here are selected snapshots, not a claim of a customer-repository commit or PR. The published experiment Results HTML contains platform-owned injected styling and is excluded from this research-source export.

Historical failed and diagnostic outputs: not confirmatory

Original paths directly under alg_hodge_27b_01/, including its shards/, roundtrip/, selected_weights.npz, failure.json, preflight and batch diagnostics, are retained as historical failed/diagnostic evidence. The first run halted because its sham failed the identity check (sham_not_identity); its apparent train/calibration measurements are not valid held-out confirmatory results. Do not mix these outputs with repaired_v2/ or interpret their uncorrected readouts as a result. The repaired experiment and final review supersede those files without overwriting the original paths. No model weights, credentials, or private library source are included. No unstated license is assigned to the published data or code.

Downloads last month
173