ken-jo/qev-data

QEV data

Training data for QEV decision models: the October 2026 vision mixes for QEV-2B and QEV-0.8B (the first release and v3), and historical snapshots for QEV 0.2.1.

October 2026 vision mix v3: QEV-2B and QEV-0.8B v3

The v3 releases of QEV-0.8B (revision bb7107f) and QEV-2B (revision 7534f3b) were trained on a revised mix, documented in vision-mix-2026-10-v3/:

• Diversified distractors: the options of 16,195 items in seven sources were regenerated with 3 to 6 options each, mixing answers from other images, perturbed values and same-image annotations. The earlier "other questions about the same image" scheme appears in 30% of these items. 60% of Mind2Web items were re-rendered with new marks and negatives.
• Targeted fixes: 707 CORD field-presence items gained receipt-text decoys, and 157 new upstream TRAIN rows (127 Hateful Memes, 30 A-OKVQA) replaced rows removed by the near-duplicate filter.
• Exclusions: 2,063 items whose images are near-duplicates (dHash distance 4 or less) of evaluation-board or internal DEV images were removed.
• Coverage mix: 51,091 items from 18 sources (49,275 trained on, 1,816 held out for selection). The manifests (config vision-2026-10-v3-coverage-manifest) list ids, gold letters, hashes and how each item was built. They contain no third-party text or images.
• Unchanged: NLVR2 ids and teacher probabilities, and the synthetic items (v3 uses 5,414 of them), are those in vision-mix-2026-10/. The 4,658 replay items are again not released.

The sources and licenses are those of the first October 2026 release below.

October 2026 vision release: QEV-2B and QEV-0.8B

QEV-2B and QEV-0.8B were trained on a new vision mix. It is documented in vision-mix-2026-10/:

• Coverage mix: 52,997 items from 18 sources (51,138 trained on, 1,859 held out for selection). For the 17 third-party sources, the manifests list each item's id, gold letter and the hashes of its options and images. They contain no third-party text or images. Build scripts are included to rebuild the items from the original train splits.
• Synthetic set: 6,000 items with images, released in full under Apache-2.0.
• NLVR2: 2,048 TRAIN, 256 DEV and 512 TEST ids, with QEV 0.2.1 teacher probabilities for TRAIN and DEV.
• Not released: 4,658 internal replay items and their targets.

The licenses differ by source; see the table in vision-mix-2026-10/README.md. FUNSD (non-commercial) and Hateful Memes (research-only) are listed by id and hash only.

The historical text and image corpora below are unchanged.

Historical corpora

Configurations overlap. Do not concatenate them or assume independent test sets.

ZIPs in corpora/ use qev-<stage>.zip and contain eligible original records, images and license notices. Extracted stage folders and original record/source IDs preserve the recorded training provenance. Viewer rows expose request/target schemas as JSON strings; parse with json.loads. IDs, group IDs, language and original split assignments are preserved.

Configuration	Rows	Train	Dev	Calibration	Test
starter-v2	8972	7008	771	714	479
policy-v3	4192	2400	448	448	896
policy-v6	6592	4800	448	448	896
policy-v8	16734	9582	1790	1786	3576
foundation-v11	19279	10841	2440	2132	3866
workflow-v12	19100	11212	2712	2344	2832
workflow-v12-cohort	23788	11212	2712	7032	2832
workflow-v12b	19100	11212	2712	2344	2832
workflow-v12c	19100	11212	2712	2344	2832
workflow-v12d	19100	11212	2712	2344	2832
workflow-v12e	19100	11212	2712	2344	2832
workflow-v12f	19100	11212	2712	2344	2832
workflow-v13g	19100	11212	2712	2344	2832
workflow-v13h	19100	11212	2712	2344	2832
workflow-v13-final	19100	11212	2712	2344	2832

Counts include repeated observations across stages; the sum is not a unique count.

Licenses and exclusions

This collection is not uniformly Apache-2.0. Read LICENSES.md, per-record source/license fields, and the original notices in licenses/. Authored data and typed-decisions use Apache-2.0; Beans/TrashNet use MIT; BANKING77 uses CC-BY-4.0; SNLI derivatives retain CC-BY-SA-4.0. Original observations were selected and converted into typed decision views. Synthetic labels are not observed outcomes; typed-decisions uses teacher agreement.

CIFAR-10 evaluation requests/images are excluded because redistribution rights are unresolved. excluded/ records references and hashes for direct upstream reconstruction. No selected training rows were removed by that exclusion.

Use and limitations

Extract one corpus ZIP under data/ in the source repository. Image paths are relative to that corpus. Manifests record original and exported record/image hashes. Group and image-byte split checks apply within each corpus, not across historical stages. They do not prove distinct physical objects or absence from Qwen pretraining. Korean generated examples are retained; personal project notes are not included. Old test groups have been inspected. Future models need untouched final populations. Full retraining from all historical intermediate checkpoints is not turnkey.

Training details.

GitHub: ken-jo/qev