Download paper/relational_physics_data/protocol.json from lamm-mit/gemma4-interpretability: direct link, hf CLI and curl.
- Browser
- Download file 3.34 kB
-
https://huggingface.co/datasets/lamm-mit/gemma4-interpretability/resolve/main/paper/relational_physics_data/protocol.json
- Command line
-
hf download hf://datasets/lamm-mit/gemma4-interpretability/paper/relational_physics_data/protocol.json
-
curl -L -o protocol.json https://huggingface.co/datasets/lamm-mit/gemma4-interpretability/resolve/main/paper/relational_physics_data/protocol.json
3.34 kB
| { | |
| "frozen_at": "2026-07-18T21:00:31.802980+00:00", | |
| "frozen_lens": { | |
| "direction": "old-law positive-minus-negative physical-outcome centroids", | |
| "large_cohort_refit": "none", | |
| "layer": 34, | |
| "position": "final_prompt", | |
| "state_normalization": "unit L2" | |
| }, | |
| "guardrails": [ | |
| "The cohort has 20 direct, 20 inverse, and 20 neutral laws.", | |
| "The 60 laws span mechanics, fracture, fatigue, transport, thermal, electrical, magnetic, optical, interfacial, microstructural, corrosion, fluid, and kinetic domains.", | |
| "Only calibration-neutral laws define center and scale.", | |
| "Direct, inverse, and validation-neutral labels never alter the normalization.", | |
| "Every law crosses two equation surfaces, two cases, two numerical directions, and two answer orders.", | |
| "The layer and hidden direction remain frozen from old laws.", | |
| "All outputs, including failures, are retained." | |
| ], | |
| "inference": { | |
| "bootstrap_resamples": 50000, | |
| "bootstrap_seed": 20260723, | |
| "permutation_seed": 20260724, | |
| "permutations": 100000 | |
| }, | |
| "inputs": { | |
| "development_manifest": { | |
| "path": "experiments/elicited-physics-abstraction-2026-07-18/prompt_manifest.json", | |
| "sha256": "4b2a691badae5519410d60250cae7fcb3f60922d00feae063ce5cc999f341894" | |
| }, | |
| "development_states": { | |
| "path": "experiments/elicited-physics-abstraction-2026-07-18/representations.npz", | |
| "sha256": "59208905855998049e3dc338dbfcb8c9fac48b935b2f946c3b33b9239d56680e" | |
| }, | |
| "prompt_manifest": { | |
| "path": "experiments/neutral-anchored-relational-physics-2026-07-18/prompt_manifest.json", | |
| "sha256": "68557163950e44d9dc8e5e38156d03ac589b30caca6615ffdce93eb3f153a946" | |
| }, | |
| "runner": { | |
| "path": "scripts/run_neutral_anchored_relational_benchmark.py", | |
| "sha256": "37a634177604686ab2a3367506649075a16c601643f4f4c62aa8aa54ce9f9274" | |
| } | |
| }, | |
| "matched_contrast": "Within each law/surface/case/answer-order cell, frozen physical score for numerical-up minus score for numerical-down.", | |
| "model": "google/gemma-4-E4B-it", | |
| "model_revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f", | |
| "neutral_normalization": { | |
| "center": "median law contrast over the 10 calibration-neutral laws", | |
| "reason": "defines an empirical physical null instead of assuming raw zero", | |
| "scale": "1.4826 times MAD over the 10 calibration-neutral laws", | |
| "validation_neutrals": "the other 10 neutral laws are excluded from normalization" | |
| }, | |
| "primary_endpoints": { | |
| "calibrated_sign_accuracy": "direct above and inverse below calibration-neutral median", | |
| "direct_vs_inverse_auc": "threshold-free, 20 versus 20 laws", | |
| "direct_vs_validation_neutral_auc": "20 versus 10 laws", | |
| "ordinal_spearman": "law sign -1/0/+1 versus neutral-normalized contrast, validation neutral only", | |
| "validation_neutral_vs_inverse_auc": "10 versus 20 laws" | |
| }, | |
| "status": "frozen-before-large-cohort-output", | |
| "study_id": "neutral-anchored-relational-physics-2026-07-18", | |
| "success_rule": { | |
| "both_surface_direct_inverse_auc_minimum": 0.85, | |
| "calibrated_direct_inverse_accuracy_minimum": 0.8, | |
| "direct_inverse_auc_minimum": 0.9, | |
| "direct_neutral_auc_minimum": 0.8, | |
| "neutral_inverse_auc_minimum": 0.8, | |
| "ordinal_permutation_p_maximum": 0.01, | |
| "ordinal_spearman_minimum": 0.6 | |
| } | |
| } | |