README.md

metadata

license: cc-by-4.0
pretty_name: Circuit-Diff paper data
tags:
  - interpretability
  - knowledge-editing
  - gemma
  - attribution-graphs
  - mechanistic-interpretability

Circuit-Diff paper data

This is the data behind the paper "Circuit-Diff: Factual Edit-based Intervention Method for Localizing Knowledge in Attribution Graphs" (Anonymous Authors, under review 2026). The code repository link appears in the paper.

Circuit-Diff diffs cross-layer-transcoder (CLT) attribution graphs before and after a MEMIT knowledge edit, ranks the resulting node changes by several candidate "diff scores," and verifies those rankings causally with targeted feature-clamping (patching) interventions. This dataset holds the pooled summary tables and one worked case-study bundle produced by that pipeline — not the raw multi-hundred-GB intervention sweeps, and not any model weights.

Provenance

Computed with Circuit-Diff v0.1.0 (the public code release; repository link in the paper) on google/gemma-2-2b (gated, subject to the Gemma Terms of Use) using the public cross-layer transcoder mntss/clt-gemma-2-2b-426k (Piotrowski & Hanna), over edits derived from CounterFact (MIT license). Attribution graphs and feature-intervention ablations were computed with circuit-tracer 0.4.1 (decoderesearch fork) (MIT license). GPU stages ran on an academic SLURM cluster.

No model weights or edited checkpoints are included in this release. Curated MEMIT edit deltas for the three case studies (enough to reconstruct each M_e = base gemma-2-2b weights + delta) are included under deltas/ — see deltas/README.md. The full ~100-edit delta set is not planned for release but exists and can be shared on request.

What's here

bundles/eiffel_paris_to_rome/   one full attribution-graph bundle (case study)
tables/setdiff-500/             appear/disappear counts + top-K overlap, 239 edits
tables/cf-sweep-24/             24-edit CounterFact ablation sweep, pooled + per-edit
tables/ablation-redo/           group/single-node patching under the correct
                                 intervention-layer protocol, incl. the null floor
tables/ablation-verification/   rank-causality verification (do diff-score
                                 rankings pick causally useful features?)
tables/score-ablation/          six diff-score rankings compared head-to-head
configs/<campaign>/             any config/edit-spec yaml or json per campaign
deltas/                         3 curated case-study MEMIT deltas — see deltas/README.md

Every table here is a pooled summary CSV already written by the campaign's own analysis code (circuit_diff.setdiff.aggregate, circuit_diff.sweep.*, circuit_diff.probpatch.report, circuit_diff.scorepatch.report). The raw per-node intervention rows (hundreds of thousands to millions of rows per campaign, tens of GB as parquet) are not included; each campaign's code in the experiments branch of the code repository regenerates them from a freshly computed bundle (cs compute + cs fetch).

bundles/eiffel_paris_to_rome/

The attribution-graph bundle behind the paper's Eiffel Tower Paris->Rome case study — the same bundle the interactive demo (cs load <bundle> --serve) reads, and the one cs print-case / print-cards / print-readouts render figures from. Same scrubbed bundle shipped with the library at circuit_diff/examples/bundles/eiffel/.

• bundle.json — run metadata (model, CLT, edit, layer settings, timings).
• gate.json — duplicate-gate membership for this bundle's nodes.
• prompt_meta.json — prompt strings and token metadata for the edit and paraphrase prompts.
• attribution/nodes.parquet, edges.parquet, errors.parquet — the attribution graph itself: per-node influence/activation values and diff scores, the edge list, and CLT reconstruction-error terms.

tables/setdiff-500/ — appear/disappear counts + null-floor overlap

Backs: the paper's set-difference + null-floor figure.

Per CounterFact MEMIT edit on Gemma-2-2b, through the frozen 426k-feature CLT, each row (setdiff_rows.csv) measures how many feature nodes an edit creates (appeared, prefix app_) and destroys (disappeared, prefix dis_) on the edit prompt, how much of a size-matched top-K set that side's diff score recovers against a graph-influence or activation reference (*_obs_infl, *_f_infl, *_obs_act, *_f_act), and a size-matched null distribution for each (*_null_infl_p2_5/med/p97_5/mean, same for _act) so the overlap reads as a claim rather than a restatement of set size. presence/ holds the same table computed under a different exclusion pass; gate-assessment/ holds a duplicate-gate probe against the cf-sweep-24 edits. Full column-by-column description and the campaign's own summary numbers (239/500 edits captured, appeared n median 20,034, disappeared n median 466 as of the last measured run) are in the campaign's own README.md / STATUS.md in the code repo's experiments branch.

tables/cf-sweep-24/ — 24-edit CounterFact ablation sweep

Backs: the 24-edit CounterFact sweep results (supplementary sweep material) and the raw edit landing/specificity table.

• edits.csv — one row per edit: whether it landed on the edit prompt and a heldout paraphrase, whether the base model's other facts stayed unchanged (specificity_ok), the old/new-answer probabilities before and after the edit, and the appeared/disappeared node counts.
• summary_pooled.csv — headline pooled stats across all 24 edits.
• movement_pooled.csv / movement_by_edit.csv — causal single-node ablation results: for each ablation dose k (disappeared: 0, -1, -2, -4, -8 times the node's own base activation; appeared: 1, 2, 4, 8 times its edited-model activation), how many labeled nodes moved the true-label vs. new-label logit/probability-rank, in which direction, and by how much (*_logit_down/up_any, *_ge_0p25, *_ge_1, *_rank_worse/better_any, n_*_rank_*, *_rank_*_max). movement_by_edit.csv adds the name (edit id) column; movement_pooled.csv pools over all 24.
• setdiff_rows.csv — same appear/disappear + overlap schema as setdiff-500/, computed for this 24-edit set (includes a dup_gate_scope and gated appeared/disappeared counts not present in the 500-edit table).

Two edits (cf943, cf14176) used capped graphs (top-32,768 feature nodes); for them "appeared"/"disappeared" means entered/left the top-k, flagged via capped / capped_detail in edits.csv. Full figure-reading notes are in the campaign's plots/FIGURE_DESCRIPTIONS.md and SWEEP_REPORT.md in the code repo.

tables/ablation-redo/ — group and single-node patching, correct protocol

Backs: the group-patching and single-node-patching figures, and the six-score comparison table (jointly with score-ablation/).

Re-runs the ablation-verification campaign's interventions under the intervention-layer protocol actually used by circuit-tracer (clamp only up to the patching end layer, not every layer) — an earlier "every layer" version of this measurement had produced an artifactual null result on the appeared side. Three experiments, each with its own summary CSVs:

• Exp 1 (inf_vs_a_signtest.csv) — does the graph-influence-based ranking beat the current champion ranking at all? Per (side, k, n_edits): how many edits the influence ranking beats on, the median gap, and whether that gap exceeds the noise floor.
• Exp 3 / group ablations (asym_group_summary.csv, pair_single_summary.csv, fivearm_summary.csv, psweep_summary.csv, single_graph_summary*.csv) — which ranking wins, and at what set size k (0.1%-25%), across arms A (the paper's diff score), INF_GRAPH, ACT_GRAPH, RND, and further baseline variants; _p95, _nofloor, _floor suffix variants repeat the measurement with/without the null-activation floor described below.
• Exp 2 / single-node ablations (scores_graph_summary*.csv, exp2/by_k*.csv, exp2/superadditivity*.csv) — feature-by-feature comparison of all eight candidate diff scores, including a superadditivity check (does patching a group beat the sum of its parts patched individually?).
• Null floor (floor_*.csv, e.g. floor_fdr.csv, floor_fdr_selected.csv, floor_coverage_exp2.csv, floor_vs_unfloored_*.csv) — a node only counts as "moved" if its activation delta clears the 99th percentile of that run's control-prompt (off-target) activation-delta pool; these tables report FDR-controlled selection and coverage under that floor, and side-by-side floored vs. unfloored results (the floor reversed some suppress-side verdicts — see the code repo's FLOOR_RESULTS.md).
• Duplicate gate (gate_*.csv, e.g. gate_share.csv, gate_coverage_exp2*.csv, gate_membership_broad.csv, gate_rerun_cost.csv) — a robustness re-analysis asking how much of each experiment's node population is a "duplicate" of an already-selected node under the duplicate gate; demoted to a robustness check because the gate can itself remove noise at rank 1, and the pre-gate numbers are treated as primary.

Column meanings and headline numbers per file are documented in the campaign's own README.md, GATE_ASSESSMENT.md, and FLOOR_RESULTS.md in the code repo's experiments branch.

tables/ablation-verification/ — rank-causality verification

Backs: the rank-causality verification result (do the diff-score rankings pick causally load-bearing features, versus activation- or random-selected baselines?).

• exp1.csv — dose-response, joint-patch medians across edits: for each (partition = appeared/disappeared, group = A/ACT/RND baseline, lam = patch strength), the median new-label recovery R_new_med, old-label recovery R_old_med, and off-target KL divergence KL_off_med.
• per_edit.csv — the same joint-patch measurement broken out per edit, plus each edit's base-vs-edited-model endpoint probabilities, the denominators used to normalize recovery scores, per-edit tau (the off-target KL normalization constant), and the golden (zero-effect sanity-check) max |Δ log-prob| for that edit.

Twelve edits, probability-space readouts, all interventions applied to the unedited model M₀. Full experimental design (why α=1, τ=0, the ε_R denominator guard, the constrained intervention-layer convention) is recorded in the campaign's README.md and REPORT.md.

tables/score-ablation/ — six-score comparison

Backs: the six-score comparison table/figure (do different candidate diff scores select causally different feature sets?).

• bands_pooled.csv — per (edit, feature group, side, arm, patch strength lam), the recovery scores R_primary/R_new/R_old, off-target KL kl, group size n_members, and which score/band selected that group (band, band_score, band_alpha — e.g. band = top/low/random 1%, 5%, 25% by a given score).
• overlap_matrix.csv — pairwise Jaccard overlap between every pair of scores' top-10% selected node sets, per side and per edit (n_a, n_b, n_inter, jaccard).

Three intervention arms: arm 1 installs candidate nodes into the base model M₀; arm 2 and arm 3 remove/suppress nodes from the edited model M_e and base model M₀ respectively. Twelve edits, edit-prompt only. Full arm/lambda-grid design and headline pooled numbers (median/p90 recovery per arm, KL quantiles) are in the campaign's README.md and REPORT.md.

configs/

Any campaign config or edit-spec file matching config*, edits*, or spec* with a .yaml/.yml/.json extension, copied per campaign (sbatch job scripts are excluded). Most campaigns select and record their edits as edits.txt (plain text, one CounterFact case id per line) or generated Python RunConfig files rather than yaml/json, so this directory may be sparse or empty for a given campaign — that reflects how the code actually stores configuration, not an omission.

deltas/

Three curated case-study MEMIT weight deltas (Eiffel Tower, Pentium II, Taj Mahal) — see deltas/README.md. No model weights or edited checkpoints are included; each delta only reconstructs M_e when applied on top of a separately downloaded google/gemma-2-2b checkpoint.

Licenses

• Tables and this dataset card: CC-BY-4.0.
• Base model: google/gemma-2-2b, used and redistributed here only in derived summary-statistic form, subject to the Gemma Terms of Use.
• CLT: mntss/clt-gemma-2-2b-426k, public, by Piotrowski & Hanna.
• Edit source: CounterFact (MIT license).
• Graph/intervention library: circuit-tracer 0.4.1, decoderesearch fork (MIT license).

Citation

If you use this data, please cite the paper (Anonymous Authors, under review 2026) and the Circuit-Diff code repository (link in the paper).