GleghornLab/esmc_tap_probing_results

Exhaustive hidden-state probing: results

Result tables of a probing study that trains a linear probe on every token-wise state of ESM Cambrian models (ESMC-300, ESMC-600, ESMC-6B; Synthyra/ESMplusplus_small, _large, _6B): the token embedding, the residual stream, the attention inputs, queries, keys and values before and after rotation, the attention context and output, and the SwiGLU internals of every block, with per-head probes on the head-split sites, and ESMFold2's structure stream (its learned 256-column mixture of every hidden state, read with the released weights of the ESMFold2 built on each model). Tasks are the 22 concept configurations of GleghornLab/protein_concept_probing and two protein-protein interaction gold standards (Synthyra/bernett_gold_ppi_2, Synthyra/bernett_gold_ppi). Code: the interpretability package of Protify (python -m protify.interpretability).

Layout

• runs/<root>/<label>/<task>/: run.json (settings, seed, model, split sizes); scores.parquet (one row per probe and condition: the full-test score and the mean and standard deviation over 7 test halves); probes.parquet (each probe's training record and validation score); elementwise.parquet (per-label scores of every probe and test half, as written by the run; it replaced the first upload's compacted ProtParam table); for the pair tasks, ranking.parquet (ROC AUC, average precision, MCC at 0.5 and balanced MCC per held-out split), intervals.parquet and bootstrap.safetensors (1,000 bootstrap draws of the test proteins).
• runs/<root>/sweeps/: settings sweeps and the selections they produced; interventions/: phosphomimetic scans, with each protein's mutant table under proteins/; latent/: principal coordinates; stores/: pooled-store manifests.
• analyses/: the trend tables and figures across models, the raw against unit-norm rule, the converged-solver check and the drawn figure sets.
• logs/: the log of every embedding, training, validation and interval step; logs/run_6b_sites/ and logs/structure_stream/ those of the two additions below.
• catalog.json (first upload) and catalog_extra.json (per-label scores, protein tables and logs, with SHA-256): every kept file with its source. Probe weights (302 GB) and pooled stores (3.4 TB) are not kept: every score is here, probes retrain from the stores, and the stores rebuild from the pinned models.

Roots: run (ESMC-300, raw states), run_norm (unit-norm), run_random* (randomly initialized ESMC-300), _s24896 and _s24897 (seed replicates; seed 24895 has no suffix), run_600_norm, run_6b_norm, run_onehot, run_randvec (baseline encodings).

Added on 2026-10-21, as new files beside the roots above, each root with a README.md:

• run_6b_sites: ESMC-6B's nine residue tasks at the 14 states inside each of its 80 blocks, with single-head probes, under run_6b_norm's code, setting, seed and controls (run_6b_norm reads its residue tasks on the residual stream only).
• <root>_struct for run_norm, run_norm_s24896, run_norm_s24897, run_random_norm, run_random_norm_s24896, run_random_norm_s24897, run_600_norm and run_6b_norm: the structure stream as one more state of that root, under its tasks, settings, seed and controls, with mean-only probes under linear_mean/.

Each probe draws its own initial weights, sees the same batches and stops on its own validation loss, so these tables join their root's per task with no probe in both; tap_probing.archive.merge_sites of Gleghorn-Lab/tap_probing does so.