moonvit-dsv4-data

Training and evaluation data for the MoonViT → DeepSeek-V4-Flash-0731 projector experiment (code: https://github.com/cyjin-yl/moonvit-deepseek-v4-glue).

All products are packed parquet with embedded PNG/JPEG image bytes (image_bytes column) — one sequential read, no small-file IO. The JSONL side is kept for content inspection; its image paths do not resolve inside this repo (images live in the parquet).

Layout

• eval_v1/packed/<name>.parquet — 5 benchmark packs (1,400 rows total): textvqa 500 / docvqa 200 / ocrbench 200 / screenspot 200 / mmmu_pro 300, each with question, answers or gt_box/gt_point, metric, image_bytes.
• eval_v1/*.jsonl + eval_v1/MANIFEST.json — same rows for inspection + per-shard sha256 of every source parquet.
• train_v1/train_mix.jsonl — 59,198 assembled short-QA rows (question/answers/image), the exact input of the mix step.
• train_v1/decontamination_report.json — per-mechanism drop counts.
• train_v1/packed/train_mix-0000X-of-00003.parquet — 59,198 rows with image_bytes (the training input; ~20 GB).
• sft_art_v1/packed/sft_art_{train,val}.parquet — full 0xSero-style art pool (71,780 train / 2,220 val, WikiArt + fashion, conversations schema), from which the mix sampled its art rows.

Mix composition (post-decontamination)

source	raw	kept	main drops
lmms-lab/textvqa train	34,602	29,252	aHash 3,999 + text 1,351
lmms-lab/DocVQA train	25,000	17,351	aHash 7,343 + text 306
showlab/ShowUI-desktop train	7,496	5,167	aHash 2,300 + text 29
art pool sample	10,000	7,428	aHash 2,572
total	~77k	59,198	23% dropped

Decontamination (three independent mechanisms, counts in the report): perceptual aHash hamming ≤ 6 vs all 1,400 eval images, exact RGB-pixel sha256, normalized question-text near-dup vs all eval questions.

Sources (all HF, shards sha256-verified against LFS oids, 0/93 mismatch)

• Eval: lmms-lab/textvqa (validation), lmms-lab/DocVQA (validation), echo840/OCRBench (test), rootsautomation/ScreenSpot (test), MMMU/MMMU_Pro standard (10 options) (test, single-image subset).
• Train: lmms-lab/textvqa (train), lmms-lab/DocVQA (train), showlab/ShowUI-desktop (train), Artificio/WikiArt_Full + benitomartin/fashion-product-images-small-384x512 (art pool).

Reproduce

python tools/prefetch_parquet.py --repo <hf-repo> --split <split> --out-dir staging/<name>
python tools/fetch_eval_data.py --dataset <name> --data-files staging/<name>/*.parquet --out-dir out/
python tools/fetch_art_data.py --wikiart-files staging/wikiart/*.parquet --fashion-files staging/fashion/*.parquet --out-dir out_art/
python tools/build_train_mix.py --source ... --eval-images eval/images --eval-jsonl eval/*.jsonl --out train_v1
python tools/pack_to_parquet.py --jsonl train_v1/train_mix.jsonl --out packed/train_mix.parquet

Load

import io, pyarrow.parquet as pq
from PIL import Image
rows = pq.read_table("train_mix-00000-of-00003.parquet").to_pylist()
img = Image.open(io.BytesIO(rows[0]["image_bytes"]))
# or via the glue repo: from tools_common import load_records; load_records(dir)

Licenses

Upstream terms apply per source dataset; most source cards do not declare a machine-readable license (checked 2026-08-04 via the HF API — MMMU/MMMU_Pro declares Apache-2.0, the other seven declare none). We redistribute derived QA rows and re-encoded images for research reproducibility under each upstream dataset's own terms; this repo grants no additional rights. Assembly code is MIT (see the glue repo).