Honey-Data-V2

[🏠 Homepage] [📖 arXiv (conference version)] [🤗 Bee models & datasets] [🍯 Honey-Data-15M] [💻 Code]

Honey-Data-V2 is the corpus of the journal version of Bee, Bee: A High-Quality Corpus and Full-Stack Suite for End-to-End Training of Fully Open MLLMs, accepted to Computational Visual Media. The conference version (ICLR 2026) released Honey-Data-15M and the Bee-8B models trained on it. The journal version scales the corpus up by reapplying the same curation pipeline, HoneyPipe, at a substantially larger scale.

Honey-Data-V2 is a multimodal supervised fine-tuning corpus of 44.3M samples comprising 100.3M QA pairs. It is stored as 19,707,852 image groups carrying 44,295,078 conversations, spread over 8 task categories and 420 subsets (5.81 TB). Long chain-of-thought covers 17.9% of the samples (7.9M, against 2.9M in Honey-Data-15M).

The released Bee-8B checkpoints, and all results reported in the paper, were trained on Honey-Data-15M. Evaluating models trained on Honey-Data-V2 is left for future work.

What is new compared with Honey-Data-15M

Honey-Data-V2 is not built from scratch. It extends Honey-Data-15M in three ways, all executed with the existing operators of DataStudio:

• Re-curation of the existing pool. Acting on community defect reports, samples whose final turn carries no answer were removed, and 492.6K grounding instructions were rewritten to state their coordinate convention up front. 1.8M machine-recaptioning samples were dropped, because they dominated by volume while adding little instruction-following signal. Image placeholders were canonicalized, and degenerate generations (unpaired reasoning tags, repetition collapse, placeholder leakage) were rejected.
• Multi-model rollout. On top of the original annotation pass, two further independent rollouts answered the same instructions with an upgraded model stack. Every verified answer is kept, so an instruction carries one, two or three independently produced answers. Over the overlapping subset, two answers to the same instruction agree verbatim in only 0.00%–0.89% of cases.
• New source incorporation. The FineVision datasets absent from Honey-Data-15M, together with Molmo2-SynMultiImageQA, ChartVerse and Aria-UI, were passed through the same pipeline rather than concatenated as-is. Sources whose answers are already determinate, such as OCR transcription and box-level grounding, were filtered but not rewritten.

One record is one image group

Unlike the usual one-row-per-conversation layout, a record here collects every conversation that refers to the same image. The image bytes are therefore stored once, no matter how many dialogues discuss it. Across the corpus this is the difference between 19,708,047 stored images and the 44,295,078 conversations that use them.

from datasets import load_dataset, get_dataset_config_names

subsets = get_dataset_config_names("HoneyDataV2/Honey-Data-V2")
ds = load_dataset("HoneyDataV2/Honey-Data-V2", subsets[0], split="train")

rec = ds[0]
rec["images"][0]                    # a PIL image
len(rec["data"])                    # how many dialogues share it
rec["data"][0]["messages"]          # [{'from': 'human', 'value': ...}, {'from': 'gpt', 'value': ...}]
rec["data"][0]["provenance"]        # which model wrote and which verified this answer

Image encoding

Images are stored as lossless WebP. Where the source was already JPEG, or where WebP came out larger than the original, the original bytes are kept untouched. WebP holds the same pixels as the PNGs it replaces in roughly 60% of the bytes. Decoding is transparent: datasets returns PIL images either way.

For RGBA images the alpha channel is preserved exactly and RGB is preserved wherever alpha > 0. PNG stores arbitrary colour under fully transparent pixels and WebP does not carry it over; no reader can observe those values.

Record fields

Each element of data:

provenance names the model used at each stage, so a subset of the corpus can be selected by annotator: question_image_consist (the image-question consistency check), answer_rewrite, reasoning_rewrite, and answer_consist (verification of the rewritten answer). A value of null means that stage did not apply — determinate tasks such as OCR transcription and box-level grounding are filtered but never rewritten.

Composition

A subset name that appears in more than one category is qualified with its category, for example Chart-CoSyn and Document-CoSyn.

Provenance of the answers

sources records which annotation pass a dialogue belongs to: Honey-Data-15M-Rollout1 is the re-curated original pool, Rollout2 and Rollout3 are the two further independent passes over the same instructions. Dialogues taken from newly incorporated community corpora carry that corpus's name instead.

Licensing information

Honey-Data-V2 aggregates many publicly available datasets, each governed by its own licence. Anyone using this corpus must observe the licence attached to each constituent dataset. To the extent we hold any rights in the curation and the generated answers, those are released under CC-BY-4.0.

Citation

If you use Honey-Data-V2, please cite the journal version:

@article{zhang2026beecvm,
  title   = {Bee: A High-Quality Corpus and Full-Stack Suite for End-to-End Training of Fully Open MLLMs},
  author  = {Yi Zhang and Bolin Ni and Xin-Sheng Chen and Heng-Rui Zhang and Yongming Rao and
             Houwen Peng and Qinglin Lu and Han Hu and Meng-Hao Guo and Shi-Min Hu},
  journal = {Computational Visual Media},
  year    = {2026},
  note    = {to appear},
}

The conference version, which introduced Honey-Data-15M and Bee-8B:

@inproceedings{zhang2026bee,
  title     = {Bee: A High-Quality Corpus and Full-Stack Suite to Unlock Advanced Fully Open MLLMs},
  author    = {Yi Zhang and Bolin Ni and Xin-Sheng Chen and Heng-Rui Zhang and Yongming Rao and
               Houwen Peng and Qinglin Lu and Han Hu and Meng-Hao Guo and Shi-Min Hu},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
}