Why this matters
Open, high‑quality supervised data for multimodal LLMs is a major bottleneck: noisy, heterogeneous sources and too few reasoning examples limit model capabilities. This release addresses that gap by reapplying a transparent curation pipeline at scale to produce a unified, provenance‑aware multimodal SFT corpus with explicit short and long chain‑of‑thought enrichment.
What Sets It Apart
- Scale with provenance: 44.3M curated samples (100.3M QA pairs) grouped into 19.7M image groups so image bytes are stored once and every answer carries model provenance — this makes subset selection by annotator/model straightforward.
- Dual‑level CoT enrichment: short and long Chain‑of‑Thought annotations (long CoT covers 17.9%, ≈7.9M samples), enabling training signals for both basic stepwise reasoning and deeper multi‑step problems.
- Re‑curation + multi‑model rollouts: the pool from Honey‑Data‑15M was cleaned (removed degenerate turns, rewrote grounding conventions, dropped excessive machine‑recaptioning), then re‑annotated with up to three independent model rollouts and a verifier, increasing answer diversity and verified quality.
- Broad multimodal coverage: 8 task categories (Caption, Chart, Document, GUI‑Grounding, General, Grounding, OCR, STEM), 420 subsets, and formats stored as lossless WebP or original bytes to preserve pixels and alpha channels.
- Practical dataset layout: one record aggregates every conversation about the same image (images stored once, dialogues in a
datalist), andprovenancefields name the model used at each curation stage so users can filter by annotator, verifier, or reasoning rewrite.
Who it's for — and trade‑offs
Great fit if you are training or evaluating multimodal SFT/RLHF pipelines, building MLLMs that need diverse grounding/vision‑language signals, or researching chain‑of‑thought supervision at scale. Useful for experiments that require provenance, multi‑answer comparisons, or chart/document/GUI grounding data. Look elsewhere or plan extra steps if you need a single uniform licence (the corpus aggregates many upstream datasets, each with its own licence), if you require human‑annotated gold answers for every sample (some answers are machine‑generated and verified), or if you need models already trained on this exact release (Bee‑8B was trained on Honey‑Data‑15M; models trained on V2 are to be evaluated in future work).
Practical notes
- Storage & size: 19,707,852 image groups, 44,295,078 conversations, total ~5.81 TB. Long CoT ≈17.9% of samples. Images use lossless WebP where appropriate and decode transparently to PIL via the
datasetsloader. - Licensing: users must respect constituent datasets' licences; curatorial artefacts and generated answers that the authors control are released under CC‑BY‑4.0.
Overall, this release offers a large, provenance‑aware multimodal SFT surface tailored for advancing fully open MLLMs, trading off unified licensing for breadth, provenance and curated reasoning signals.
