Download paper/graph_data/README.md from lamm-mit/gemma4-interpretability: direct link, hf CLI and curl.
- Browser
- Download file 3.02 kB
-
https://huggingface.co/datasets/lamm-mit/gemma4-interpretability/resolve/main/paper/graph_data/README.md
- Command line
-
hf download hf://datasets/lamm-mit/gemma4-interpretability/paper/graph_data/README.md
-
curl -L -o README.md https://huggingface.co/datasets/lamm-mit/gemma4-interpretability/resolve/main/paper/graph_data/README.md
Graph-analysis Supplementary Information data
This directory is the portable, machine-readable record behind the graph
sections of the main paper and Supplementary Information. It is copied from
the fingerprinted analysis at
experiments/graph-topology-rigorous-2026-07-17/ so that the complete
paper/ directory can be moved to Overleaf or archived independently.
The graph study is a post-hoc analysis of previously generated and inspected Gemma states. The primary graph protocol was frozen before graph outputs were calculated. Later falsification and ranking analyses were frozen in separate amendments before their corresponding outputs. The graph evidence is representational, not causal, and the signed-relation graph is stratified by the supplied mechanism family.
Prompts and primary edges
signed-relation-prompt-manifest.json: all 72 exact prompts, mechanism and material-case identifiers, surface variants, physical directions, expected answers, and displayed answer orders.primary_jacobian_edges_with_ids.csv: all 144 primary frozen-band Jacobian edges with complete prompt IDs, case IDs, physical directions, expected outcomes, similarities, correctness, first-valid ranks, and pairwise AUCs.all-method-selected-edges.csv: all 1,440 selected-edge rows across matched state, lexical, answer-order, oracle, and layer-band configurations.all-method-candidate-ranks.csv: all 1,440 corresponding candidate-ranking rows.
The Supplementary Information prints every prompt and every primary edge in human-readable form. Regenerate those LaTeX inputs from the repository root:
python scripts/build_graph_si_inventory.py
Layer, robustness, and negative-control data
relation-layer-metrics.csv: signed-relation edge precision by layer and representation.mechanism-layer-metrics.csv: 50-prompt cross-phrasing mechanism graph by layer and representation.mechanism-residual-layer-metrics.csv: mechanism graph after prompt-only similarity residualization.mechanism-hard-negatives.csv: all frozen lexical-competitor comparisons.target-free-word-enrichment.csv: all 523 target-free word-network tests; zero survives BH FDR 0.05.statistics.json: complete numerical output, including blocked permutations, the 46,656-assignment exact null, family results, ranking, leave-one-family-out checks, edge stability, and negative controls.validation.json: independent 11-check audit record.
Frozen analysis record
protocol.json: primary graph protocol.protocol-amendment-v1.json: answer-order and exact structured-null falsification plan.protocol-amendment-v2.json: ranking and robustness plan.protocol-amendment-v2-correction.json: preserved implementation correction.
The manuscript records SHA-256 fingerprints for these files. The authoritative analysis and audit commands are:
python scripts/analyze_graph_topology_rigorous.py
python scripts/audit_graph_topology_rigorous.py