qs-xiao/submission-2026-anon

DV-ViT Measurements

Per-input measurement records from the DV-ViT evaluation protocol on token-reduced Vision Transformers (NeurIPS 2026 Evaluations & Datasets Track submission, currently anonymized for double-blind review).

Dataset summary

This release packages the per-input measurement records produced by running five pinned sub-protocols against trained Vision Transformers and four token-reduction methods. The protocol — not the records — is the principal contribution of the parent paper. The records are released so that the protocol's verdict criteria, soundness inclusions, and audit-pair findings can be verified, extended, and applied to new methods or backbones.

Five RecordSets, 2934 records total.

RecordSet	Records	Shape	What it measures
m1-headline	500	wide (1 row / input)	CROWN bound widths vs PGD-fast and PGD-strong widths on a 2-layer auto_LiRPA-compatible ViT under native LayerNorm at ε=2/255 (CIFAR-10)
m1-bn-ablation	500	wide	Same protocol on the architecture-matched BatchNorm-substituted ViT (paired with m1-headline by image_idx)
m2-route-stability	50	long (1 row / input × method)	Bisection-resolved smallest perturbation δ̂ at which the reduction route flips, on 5 methods × 10 stratified ImageNet-val inputs
m3-discrepancy-ceiling	650	long (1 row / input × method × eps × pgd_budget)	PGD-maximized L2 discrepancy between dense and reduced models at 7 ε under the original 50×3 PGD budget for ToMe / DynamicViT / EViT (210), plus 3 ε under the 200×10 high-budget for those three from an earlier vintage (90), plus the full 7-ε grid at the 200×10 high-budget for those three from a subsequent full-grid rerun vintage (210, cross-vintage protocol parity at deployment ε confirmed within bootstrap noise), plus the full 7-ε grid at 200×10 high-budget for DiffRate and ATS (140). Total 650 records; see the pgd_budget and source_experiment columns
m4-mechanism-arc	1234	long, partitioned by arc	Mechanism-arc records spanning a register-token isolation arc, a supervised-vs-self-supervised arc, a reduction-rate sweep arc, a per-patch merge-margin probe arc, a slot-identity strict-test arc, and a flip-localization arc

Supported tasks

This is a measurement dataset, not a learning dataset. It does not have a target label, classification task, or generation prompt. The intended uses are:

• Reproducing any per-input number cited in the parent paper, by reading the corresponding row of the corresponding RecordSet.
• Auditing the soundness inclusions on the slack measurement (M1) (CROWN ⊇ PGD-fast ⊇ MC, and nominal-in-CROWN) — the three soundness flags are stored explicitly per input.
• Comparing verdict criteria across methods or protocol variants without rerunning the experiments — e.g. computing alternative aggregations of the bootstrap CIs at different eps, or stratifying by is_calibration.
• Extending the protocol to a new token-reduction method by following the experiments/external-reuse-*/ adapter pattern (see in `).

Languages

The records are language-neutral. Dataset metadata, field names, and documentation are in English.

Dataset structure

Each RecordSet is shipped in three formats:

• <recordset>.parquet (primary; snappy-compressed)
• <recordset>.jsonl (line-oriented JSON, one record per line)
• <recordset>.csv (lowest-common-denominator CSV; nested fields encoded as JSON strings)

The Croissant 1.0 metadata is in croissant.json. A field-level provenance file is in provenance.csv (one row per (record_set, field_name) mapping to the source experiment, governing, and dtype).

m1-headline (500 records, 41 fields)

Per-input record from the M1 LayerNorm headline sweep:

• i, image_idx, label, class_name, pred, pred_class, is_calibration
• crown_*: per-class lower/upper bounds, widths, and avg width from CROWN
• pgd_strong_*: same shape, from 200-step × 10-restart PGD (only on the 20 calibration inputs; null elsewhere)
• pgd_fast_*: same shape, from 50-step × 3-restart PGD (all 500 inputs)
• mc_*: same shape, from 100k uniform-noise Monte Carlo samples
• slack_factor_fast, slack_factor_strong: CROWN avg width / PGD avg width
• soundness_nominal_in_crown, soundness_pgd_fast_subset_crown, soundness_mc_subset_pgd_fast
• iter_wall_seconds, iter_rss_peak_mb, iter_gpu_peak_mb

m1-bn-ablation (500 records, 41 fields)

Same field schema as m1-headline; protocol-matched run on the BatchNorm-substituted architecture. Use image_idx as the join key to pair with the LN headline for the ratio analysis.

m2-route-stability (50 records, 22 fields)

One row per (method, trial):

• method ∈ {ToMe, DynamicViT, EViT, DiffRate, ATS}
• delta_hat_pixel: the headline (in pixel/255 units)
• first_flip_eps_pixel, last_noflip_eps_pixel: pre-bisection grid
• bisection_iters, bisection_bracket_pixel, bisection_delta_hat_pixel, bisection_trace
• grid_records: JSON-encoded per-eps grid points
• Per-trial telemetry: trial_wall_s, trial_gpu_peak_gb, trial_rss_gb

DiffRate's records were added by the first external-reuse adapter demonstration; ATS's records were added by the second external-reuse adapter (a parameter-free differentiable sampler).

m3-discrepancy-ceiling (650 records, 21 fields)

One row per (method, trial, eps_pixel, pgd_budget, source_experiment):

• method ∈ {ToMe, DynamicViT, EViT, DiffRate, ATS}
• pgd_budget ∈ {original-50x3, high-200x10}; do not average across budgets
• eps_pixel ∈ {0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0} for original-50x3 and now (post the full-grid rerun) for high-200x10 on all five methods
• source_experiment: identifies vintage. For ToMe/DynamicViT/EViT at high-200x10, two vintages are preserved: an earlier (3-eps deployment-aligned) and a subsequent (7-eps full-grid). Cross-vintage protocol parity at deployment ε ∈ {8, 16, 32}/255 is confirmed within bootstrap noise (DynamicViT/EViT bit-identical medians; ToMe drift ≤ 1%). Group by (method, pgd_budget, eps_pixel, source_experiment) to dedupe; or filter to a single vintage with source_experiment.str.contains("the subsequent vintage") for the full-grid 7-ε analysis
• d_worst: worst-case L2 discrepancy across PGD restarts
• d_0: clean L2 discrepancy
• pred_dense_top1, pred_reduced_top1, top1_agree
• per_restart: JSON-encoded per-restart trajectory

Bootstrap CIs at eps=8/255 for headline d_eps cited in the paper are computed from this RecordSet's high-200x10 rows; see paper/shared/figures/bootstrap_m4.py in the parent repo for the reproducible aggregation script.

m4-mechanism-arc (1234 records, 29 fields)

Heterogeneous multi-arc RecordSet. Use the arc and grain columns to filter:

• arc=register-isolation, grain=per-image (225 records): register-token isolation on DINOv2 (75 inputs × 3 variants)
• arc=supervised-vs-self-supervised, grain=per-image (150 records): DeiT-S supervised-vs-DINO-v1 (75 × 2)
• arc=reduction-rate-sweep, grain=per-image (600 records): DINOv2 reduction-rate sweep (75 × 8)
• arc=merge-margin, grain=per-sentinel (102 records): ToMe merge-margin probe with stability-probe cross-check
• arc=slot-identity, grain=per-sentinel (82 records): slot-identity strict-test rerun
• arc=flip-localization, grain=per-image-multi-method (flip-localization arc, 75 records): multi-method flip localization

Arc-specific structures (per_block, flip_localization_per_layer, stability_probe_cross_check) are JSON-encoded as *_json columns to preserve heterogeneity. Use the parquet file as the primary access; the JSONL preserves the same JSON encoding line by line.

Data splits

This is a per-input measurement dataset; there is no train/val/test split in the conventional sense. All records are released as a single train split per HuggingFace convention. Use is_calibration (M1) to filter the 20 calibration inputs out of the 500-input headline sweep.

Field provenance

Each field's provenance is documented in provenance.csv:

record_set, field_name, dtype, governing_adr_for_record_set,
source_experiment(s), description

The governing_adr_for_record_set column points into of the parent repo. Thesource_experiment(s)column lists theexperiments//directory(ies) whoseresults.jsonl` produced the records.

Source data

Initial data collection and normalization

Records are programmatic outputs of the experiment.py scripts in each source experiment directory. No human annotation. Source images are from ImageNet-val, Hard ImageNet (Moayeri et al. 2022), and CIFAR-10-test. The assembly script assemble.py reads the source results.jsonl files, normalizes field names across method variants (e.g. flipped_layers → flipped_blocks), and writes the parquet/JSONL/CSV artifacts. The script is deterministic and idempotent; rerunning it from the same source results.jsonl files reproduces the artifacts byte-for-byte.

Who are the source data producers?

The measurements are computed by the parent paper's authors (currently anonymized for double-blind review). Upstream data attribution:

• CIFAR-10: Krizhevsky et al., 2009.
• ImageNet (1000 classes): Deng et al., 2009. The records do not redistribute the ImageNet images themselves; only programmatic measurements computed against them.
• Hard ImageNet (15-class spurious-cue subset): Moayeri et al., 2022, CC-0 license.

Annotations

None. All values are programmatic measurements.

Considerations for using the data

Social impact

This is a measurement dataset for neural-network verification research. It does not contain personal data, biometrics, or sensitive content. Its direct social impact is bounded by its niche: tooling for the verification research community.

Discussion of biases

The records reflect the protocol's choices, not properties of "ViTs in general":

• M1 records derive from a 2-layer ViT specifically constructed to be auto_LiRPA-compatible. Production ViTs are deeper, wider, and trained differently; M1 conclusions do not transfer.
• The route-stability measurement (M2) records use n=10 stratified ImageNet-val inputs per method. This is commensurate with verifier-benchmark practice but small for absolute claim precision; bootstrap CIs are reported in the parent paper.
• The mechanism analysis (M4) records operate on Hard ImageNet's 15-class spurious-cue subset (75 inputs), and inherit that benchmark's documented properties (foreground/ background mask provenance, class imbalance, etc.).

Other known limitations

• M2 combines two PGD budgets. The pgd_budget column distinguishes them; do not average across original-50x3 and high-200x10. The paper headlines use the high-budget rows.
• Cross-pod tolerance: the protocol pins a 3× tolerance band on M1 cross-pod reproduction. Absolute slack-factor values within [5.54×10⁸, 1.66×10⁹] (the M1 LN-headline median 5.54×10⁸ and its 3× upper bound) on Ampere-class hardware constitute a successful reproduction.
• Reproducibility anchors live in the source-experiment manifests. Each released record's source_experiment field points to experiments/<slug>/manifest.json in the parent repo. The manifests carry (where available): engine SHAs auto_lirpa_sha + alpha_beta_crown_sha; conda env name alpha-beta-crown (env spec at docs/baselines/3090-original/env_strict.yml); gpu_name + torch_version + cuda_version; session_id resolving to a session-log block with the session-end git SHA; and a precursor- pointer (precursor_decisions list or hypothesis_pre_registered text). Coverage is uneven across the corpus:
  • Engine SHAs: present on the original four-method M2 + M2 + mechanism-probe extension arc; the ATS manifest (covering the +80 ATS records) carries ats_sha only and assumes engine continuity with the surrounding session window — see experiments/external-reuse-ats/GATES_2_3_SUMMARY.md for the Gate-3' verification that ATS source code does not modify the engine.
  • session_id was added to the experiment-template wrapper partway through the project; M1 LN-headline + M1 BN-ablation manifests (1,000 records) lack it. Their session-log window is recoverable from the broader docs/ chronology and the precursor- list (LN-vs-BN ablation, training protocol).
  • environment.yml hash is not captured. Manifests record env name alpha-beta-crown only; the canonical env-spec snapshot lives at docs/baselines/3090-original/env_strict.yml (which pins the conda version manifest used across the project).
• Project-side repo_sha is a secondary lookup key, not the primary reproducibility pin. It was added to the experiment-template wrapper at post-Checkpoint-3. The released records additionally carry a repo_sha_validity column distinguishing three states:
  • 497 records (18%) — valid: full 40-char SHA resolves directly as a git object. Coverage: m2-route-stability (40), m3-discrepancy-ceiling original-budget arc (280), m4-mechanism-arc merge-margin + flip-localization (177).
  • 670 records (25%) — resolved-from-prefix: abbreviated / synthetic-zero-padded SHA whose 7-char prefix resolves unambiguously to a real commit; the assembly script writes the canonical 40-char form. Coverage: m1-headline (500) + m3-discrepancy-ceiling high-budget arc (160) + m2-route-stability ATS slice (10).
  • 1557 records (57%) — unresolved: source manifests pre-date the post-Checkpoint-3 wrapper change and recorded a null git_sha field. The manifest-side anchors above remain intact for these records (engine SHAs + pointers + GPU / torch / cuda); only the project-side commit pointer is null. Coverage: m1-bn-ablation (500) + m4-mechanism-arc {the register-isolation arc, the SSL-vs-supervised arc, the r-sweep arc extension, slot-identity-rerun} = 1057.
• A camera-ready sweep is documented as a presentation-cleanliness follow-up that (i) backfills repo_sha on pre- records where the session-log window narrows the search to a single commit, and (ii) materializes the manifest-side anchors directly into the released JSONL rows so external users do not have to follow the source_experiment indirection. This does not affect measurement values; it improves anchor accessibility.
• The DiffRate records (added per) use a slightly different autograd-attackable merge metric than the strict DiffRate operator; see experiments/external-reuse-diffrate/README.md for the honest deviation disclosure.
• The ATS records (added per) operate the released AdaptiveTokenSampler in a two-mode regime: drop_tokens=True for the inference forward, and drop_tokens=False for backward-pass protocols (PGD against route-flip), because the released scatter(reduce='add') compaction in drop_tokens=True is non-differentiable in PyTorch 2.8+. The two modes produce identical surviving-token sets on clean inputs; the regime is documented at experiments/external-reuse-ats/GATES_2_3_SUMMARY.md.

Discouraged uses

• Treating the median CROWN/PGD slack on M1 records as evidence of a property of any deployed ViT.
• Treating M2/M3 sub-pixel radii or discrepancy ceilings as direct attack budgets for any production-scale ViT.
• Treating absences in this corpus (methods/eps/backbones we did not include) as evidence of method-level invariants.

Licensing

• Measurement records (this dataset): CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/)
• Code in the parent repository: Apache-2.0 (matches upstream auto_LiRPA / α,β-CROWN engine licensing)
• Hard ImageNet pixel masks referenced in m4-mechanism-arc records: inherit CC-0 from Moayeri et al. 2022.

Versioning

Semantic versioning. Current release: 1.0.0. Each measurement value is anchored to its source-experiment manifest at experiments/<slug>/manifest.json in the parent repo (the record's source_experiment field is the lookup key); the manifest carries the engine SHAs, environment, hardware, and protocol-pinning pointers per the "Other known limitations" coverage breakdown. The record-level repo_sha field is a secondary lookup key, not the primary reproducibility pin. Patch versions for metadata-only fixes; minor versions for any measurement update (always recorded in the parent paper's ledger).

Citation

@misc{anonymous2026dvvit,
  title = {Per-input measurement records for the DV-ViT evaluation protocol},
  author = {Anonymous (NeurIPS 2026 ED-track double-blind submission)},
  year = {2026},
  note = {Anonymized; the public release will replace this notice with the author block at camera-ready.}
}

Reproducibility

The release was assembled via paper/shared/release/assemble.py from the parent repository, with deterministic outputs given the same source experiments/ directories. To reproduce the artifacts from the source experiments:

# from the parent repository
python paper/shared/release/assemble.py
python paper/shared/release/generate_provenance.py
python paper/shared/release/generate_croissant.py
python paper/shared/release/validate_croissant.py

Contributions

We thank the upstream authors of α,β-CROWN, auto_LiRPA, ToMe, DynamicViT, EViT, DiffRate, DINOv2, Hard ImageNet, and AutoAttack — all named in the parent paper's literature notes ().