Ahsanz/e2-counterfactual-geometry

E2: counterfactual geometry items with certified edits

Programmatically generated geometry and chart questions where every item ships with paired counterfactual images: edits that MUST change the answer (necessary) and edits that MUST NOT (irrelevant), each certified by re-solving the scene. The benchmark measures whether a vision-language model's answer is grounded in the figure: necessity (does the answer follow a change that matters?) and invariance (does it ignore a change that does not?), rather than accuracy alone.

Generated by the benchmark/e2 package of the accompanying repository; the code, templates, filters and harness are the source of truth for every field here. This repository holds the built datasets so the code repository does not carry ~100k images.

Datasets

folder	items	templates	purpose
v2/	6618	50 (10 Geo3K-style + 40 flavoured from MathVerse, MathVision, DynaMath)	the paper's v2: three benchmarks in one manifest, mixed (2962), mode_g (2937), mode_l (3681)
v1/	1477	10 Geo3K-style	frozen v1 (e2-v1-schema-freeze)
v2_pilot/	900	30	gate D2' pilot A (30 items per template, pre-revision)
v2_pilot_b/	780	26	gate D2' pilot B (stretch tier + six revised templates)
pilot/, pilot_v1_withT5/	~50	5-6	the original v0 pilots

Each folder holds the dataset's text artefacts as plain files and its payload directories as one uncompressed tar each (the Hub rate-limits per-file operations, and these are ~100k small files):

<dataset>/
  manifest.json        # item index with benchmark membership, style_family, pool sizes, render hash
  FUNNEL.md, F4_BALANCE.md, FILTERS_F4_F5.json, REVIEW*.md   # build funnel, filter reports, review sheets
  items.tar   -> items/<item_id>.json    # question, gold answer, evidence, interventions with new answers, filter verdicts
  images.tar  -> images/<item_id>_clean.png, <item_id>_<edit_id>.png   # 532 x 532 = 19 x 19 Qwen2.5-VL visual tokens
  masks.tar   -> masks/<item_id>_<edit_id>.png                          # pixel mask of each edit
  leaky.tar   -> leaky/<item_id>.json    # items removed by the blind-text filter F5 (kept for the leakage analysis, not scored)

scripts/e2_hf_dataset.py download <name> in the code repository fetches a folder and extracts the tars in place, which yields exactly the layout the harness reads.

Item schema (0.2)

item_id, template_id, evidence_mode (G|L), style_family, source_subject, perception_demand, benchmarks [mixed|mode_g|mode_l], question, answer {value, type, unit, tolerance}, image {path, render_config_hash}, evidence {elements, patch_indices_qwen25vl}, interventions [{edit_id, edit_type (necessary|irrelevant), edit_op, target_element, new_answer, image_path, edit_mask_path, changed_patch_indices_qwen25vl, visibility, certification, filters.f5b_leaked}], filters.

Consumers must exclude edits with filters.f5b_leaked = true from necessity metrics: their new answer is guessable from the question text alone.

Filters applied

F0 render lint, F1 solver certification of every edit, F2 visibility (each edit changes at least one visual token inside its mask and nothing outside), F3 dedupe on quantized parameters, F4 answer-prior balance, F5 blind text-only model (Qwen2.5-VL-3B, image withheld, floor-aware; 1.21% of v2 items removed to leaky/), F5b the same on new answers (124 edits marked). No item was ever filtered on a sighted model's behaviour.

Use

python scripts/e2_hf_dataset.py download v2      # snapshot_download + extract the four tars
python -m benchmark.e2.harness.runner --manifest benchmark/e2/datasets/v2/manifest.json ...

Citation

Paper in preparation (2026). Until then cite the repository.