mohit67890/imajev-bench

ImajevBench v2.0-lite (preview)

A benchmark for typed decisions made from a photo, a written rule, or both. A system receives the evidence (0–2 images and a state: an order, a rule, a form, a claim) and one typed question, and must return exactly one of: yes/no, a listed choice, an integer level, or Unknown when the evidence does not determine the answer. It checks four things: whether the decision is correct, whether the system abstains exactly when it should, whether the answer changes when and only when the evidence changes (contrast sets), and, for systems that return probabilities, whether they are calibrated.

It exists because the public typed-decision benchmarks (JevBench and its relatives) are text-only. Deciding from a menu board, a parking sign, a price tag, a nutrition panel or a shelf against a rule is a different task, and it is the task small local models are being asked to do on device.

Status: preview. Read the caveats before quoting any number.

• 533 items: 75 text, 223 visual, 235 joint (image + state), over 317 images from 98 scenes; 45 references are Unknown.
• Splits by evidence cluster: dev 173, calibration 81, test 279 (89 test clusters).
• All images are AI-generated (gpt-image-2.5-flare: 385 references, gemini-3.1-flash-image: 73), from structured scene specifications; answers are known by construction. Each image was kept only when two checker models confirmed every stated fact. There are no real photographs in this version, so results say nothing direct about real-world photos.
• Labels are model-audited, not human-audited. 138 items (a seeded 15% cluster sample, every occlusion-based Unknown, every flagged item) were answered blind by two model families outside the generator, checker and leaderboard roles (Claude Opus 5.5, Kimi K2.5). Result: 1 label error in 138 (0.7%, upper 95% bound 4.0%); that item is quarantined and not scored.
• Release gates not met (see DATASHEET.md, pilot-size targets): statistical power (89 test clusters against about 99 needed to detect a 10-point paired difference); construction lint (one near-duplicate flag, inspected and adjudicated as a false positive); image necessity, which the release check could not evaluate from a test-split-only no-image run (on the test split, imajev-2b without images abstains on 258 of 279 items).
• Frontier API models score about 99% here; this version does not separate them and is disclosed as a sanity check for them. The intended subjects are small and local models.

Scoring rules

• Unknown is correct only when the reference is Unknown. Answering Unknown on an item that has an answer is wrong; giving an answer on an item whose reference is Unknown is wrong; malformed replies and timeouts are wrong. Correct-Unknown and false-abstention rates are reported beside accuracy and do not change it.
• Chance-corrected accuracy = (accuracy − chance) / (1 − chance). Chance on the test split is 26.5%, a uniform guess over each item's allowed answers including Unknown; the per-track option-count table is in LEADERBOARD.md.
• Calibration (ECE, Brier) is reported for systems that return probabilities (direct option scoring).
• Intervals. 95% intervals are bootstraps over the 89 evidence clusters: about ±6 points for systems scoring 60–80%. Differences smaller than that are not established; the paired cluster test decides comparisons.

Files

• records/records-public.jsonl: all 533 records. Dev and calibration rows carry gold; test rows have gold withheld (gold_withheld: true) and their construction truth removed. Test labels are held by the maintainers (see Submitting).
• assets/: the 317 images, content-addressed (assets/<sha256-prefix>.jpg), referenced by relative path from the records.
• LEADERBOARD.md: test-split results with interface and reasoning setting per row.
• DATASHEET.md: composition, release gates, collection and licensing, annotation, intended use, maintenance.
• harness/: pointer to the code (src/imajev_bench in https://github.com/mohit67890/imajev).
• runs/: the released imajev models' test-split runs (predictions.jsonl, manifest.json, completion.json, raw.jsonl, score), the folders behind their leaderboard rows.

Dataset version v2.0-lite; records/records-public.jsonl sha256 8551501a51853e91689d5970a2a7f7218b1de48729b9743ba620d2ca32af6ba8. Scores are comparable only within one data version.

Record shape

{"id": "scene-03-0000-base-q0", "group_id": "grp-…", "track": "joint", "family": "multi_step_rule", "split": "dev",
 "images": [{"path": "assets/51d4….jpg", "sha256": "…"}],
 "request": {"schema_version": "1.0", "state": {"order": "1 × Iced Tea, 1 × Cola", "rule": "Approve the order if and only if …"},
             "fields": [{"id": "decision", "question": "Is the order approved under the rule?", "type": "boolean"}]},
 "gold": true, "annotation_status": "reviewed", "provenance": {"contrast": {"set_id": "…", "role": "original"}, "…": "…"}}

Families: answerability, attribute_state, comparison, counting, date_arithmetic, missing_conflicting_evidence, multi_clause_rule, multi_step_rule, numerical_reconciliation, policy_precedence, rule_exception, spatial_counting, spatial_relation, text_reading, threshold_rule (see docs/imajev-bench-v2-families.md in the code repository).

Running a model

Two interfaces, reported in separate leaderboard groups because they are not comparable:

1. Direct option scoring (open-weights models): native template, profile benchmark-neutral, all option orders averaged, Unknown appended as a candidate, uncalibrated unless fitted on the calibration split only.

   # in the imajev repository, with the dataset folder next to it
   PYTHONPATH=src python scripts/imajev_bench/run_local_v2.py --model imajev2b-v21 \
     --records <dataset>/records/records-public.jsonl --split dev --allow-draft --condition full \
     --output runs/imajev2b-dev --gpu-coordinated
   python -m imajev_bench score --records <dataset>/records/records-public.jsonl --split dev --predictions runs/imajev2b-dev/predictions.jsonl --output runs/imajev2b-dev-score
   

   Required runs per model: full; no_image on visual and joint items (image-necessity control); no_state on joint items.
2. Structured generation (API models): the model returns JSON matching the field type; malformed or out-of-domain outputs count as errors, never as abstentions, and the parse-failure rate is reported. python -m imajev_bench api-run --provider … --model … --records … --split dev --output …

Every entry states model revision, precision, prompt profile, visual token budget, interface and reasoning setting. Timing is a secondary column measured on an idle machine.

Submitting

Test images and requests are public; test labels are not. Run the harness on the test split (--split test) and open an issue in the code repository attaching the run folder (predictions.jsonl, manifest.json, completion.json; the receipts bind the predictions to the exact records and settings). Up to 3 scored submissions per team per month. Test labels are released after 12 months or when v3 ships, and the split then retires to dev.

Hidden test. The maintainers also hold a private test set that is never published (202 items in 66 evidence clusters, built the same way from new scenes, model-audited). Send model weights, which the maintainers run offline, or an API endpoint, whose owner can see the items. Only aggregate metrics and the hidden set's hash are published. Policy: docs/imajev-bench-hidden-test.md in the code repository.

Leaderboard

See LEADERBOARD.md. Test split, 279 items; the two interfaces are ranked separately.

• Direct option scoring: imajev-4b 82.4% [0.77, 0.89], imajev-9b 82.1% [0.76, 0.88], Qwen3.5-9B base 76.7%, imajev-2b 71.7% [0.65, 0.78], Qwen3.5-4B base 70.6%, Qwen3.5-2B base 60.2%. The previous versions of the imajev models (previous version) are also in the table.
• Structured generation: Gemini 3.1 Pro 99.6%, GPT-5.6 Luna 99.3%, GPT-5.4 (medium reasoning) 98.9%, Gemini 3.8 Flash 98.2%, Grok 4.3 91.4%; open models 58.8–74.6%.
• Controls: without images, imajev-2b abstains on 258 of 279 items. Without the written state, it falls from 76 to 46 of 122 joint items.

Licence

Records, questions, labels and scene specifications: CC BY 4.0. Generated images: CC0-1.0 (as recorded per image in provenance.image_sources[].license). Harness and scorer: Apache-2.0 (code repository). Please cite as "ImajevBench v2.0-lite (preview), 2026, https://github.com/mohit67890/imajev".

Conflict of interest

The imajev models' developer built and maintains this benchmark. imajev results here are self-reported; the external reviewer required by the benchmark plan has not yet signed off this version.