MMProLong-Data-Public

This is a newly constructed public corpus associated with MMProLong, not the original training data used in the paper. Due to ByteDance's internal data governance and management requirements, we cannot open-source the paper's original training corpus. This release is constructed separately from publicly accessible documents. It supports research on long-context document understanding; it does not establish reproduction of the paper's reported results.

Current production uses GPT-5.6, recorded as gpt-5.6-sol. The release also retains historical GPT-5.5 examples, with actual model identifiers on every QA row. See generation provenance for current counts and the model-specific loading example below.

Public snapshot (2026-11-09): 34,212 QA examples, 5,458 source PDFs, 128,188 unique page images, and 2,079,092,954 estimated training tokens (approximately 2.079B).

If you use the dataset, annotations, or synthesis methodology, please cite the MMProLong paper. The BibTeX entry is provided in Citation.

Quick links

• Relationship to the paper
• Dataset details and statistics
• Classification and annotation schema
• Generation provenance
• Download and use
• Files and formats
• Quality checks and limitations
• Reproduce the statistics
• Citation

Relationship to the paper

The paper is Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context, NeurIPS 2026. Its code is available in the MMProLong GitHub repository.

Aspect	Original paper	This public release
Data availability	Original corpus cannot be released under ByteDance's internal data governance requirements	Separately constructed public-document corpus
Document sources	A multi-source pool of 1,537,504 PDFs, including papers, books, and manuals	Admitted public research papers; see measured source counts below
Generation model	Seed 2.0 for the paper's VQA synthesis	Current pipeline: GPT-5.6-sol; retained historical GPT-5.5 examples are labeled
VQA construction	Generate from coherent 8-15-page segments, then restore full-document context	Full-document generation/review, explicit local/distributed evidence constraints, and unanswerable examples
Default VQA page range	32-50 pages	Observed range: 14-50 pages
Supervision	Three VQA tasks; OCR transcription studied separately	Five QA types; no OCR-transcription training split
Experiment claims	Results and ablations reported in the paper	No claim that this snapshot produced those results

The paper's default VQA study reports 177,774 examples and approximately 15.16B tokens across three task-specific pools. Its 5B-token training budget is per experimental run. Those quantities describe the paper's data and experiments, not this public snapshot. The manuscript uses binary prefixes (K = 2^10, B = 2^30); quoted paper statistics retain that convention. This public card uses decimal K = 1,000 and B = 1,000,000,000, with exact token counts available. The >=100Ki statistic uses 102,400 tokens to match the paper's 100K threshold.

Dataset details and statistics

All counts below describe admitted QA examples. Rejected candidates, held documents, and uncommitted production outputs are excluded. A unique PDF is counted once in document statistics; its context is counted again for each QA in training-sample statistics.

Corpus scale

Statistic	Public snapshot value
Admitted unique PDFs	5,458
Document family	Public research papers
Unique rendered page images	128,188
Document pages: mean / median	23.49 / 21
Document pages: range / 5th-95th percentile	14-50 / 14-41
Admitted QA examples	34,212
Estimated training tokens	2,079,092,954 (2.079B)
Per-QA tokens: mean / median	60.771K / 53.469K
Observed token range	33,547-125,503
At least 100,000 tokens	2,602 (7.61%)
At least 100Ki = 102,400 tokens (paper threshold)	2,186 (6.39%)
Page references across all QA contexts	818,295

Document sources

Source	Unique documents
arxiv	4,386
elife	17
plos	1,055

Tokens are the stored Qwen2.5-VL serialized-sample estimate, including full page-image context and QA/prompt text. A PDF used by several questions contributes its context repeatedly to the training-token total. This is not a count of unique-text tokens or billed API tokens. Page images are physically stored once per document.

Task-level statistics

Task	QA	Avg. pages	Avg. K tokens	B tokens	Pages	K-token range	>=100Ki
single	12,388	23.80	60.44	0.749	14-50	33.80-125.13	6.34%
multi_local	6,134	23.93	60.88	0.373	14-50	33.85-125.50	6.33%
multi_distributed	6,172	23.87	60.68	0.374	14-50	33.84-125.29	6.40%
reasoning	6,175	23.88	60.64	0.374	14-50	33.55-125.11	6.09%
unanswerable	3,343	24.51	62.23	0.208	14-50	33.79-125.09	7.24%
Total	34,212	23.92	60.77	2.079	14-50	33.55-125.50	6.39%

K = 1,000, B = 1,000,000,000; 100Ki = 102,400 tokens. Average pages in this table are QA-weighted, whereas the corpus-scale document average counts each PDF once. The two multi-page rows may be combined for comparison with the paper's broader extraction category. Unanswerable QA is an additional public-data task. Exact values, quantiles, and the combined task grouping are in metadata/statistics.json.

Primary-domain distribution

Primary domain	Documents	Document share	QA	QA share
Physics	1,321	24.20%	9,282	27.13%
Mathematics	1,223	22.41%	8,802	25.73%
Computer Science	1,174	21.51%	9,019	26.36%
Engineering	368	6.74%	2,673	7.81%
Biology	301	5.51%	723	2.11%
Medicine	294	5.39%	920	2.69%
Economics	289	5.29%	1,020	2.98%
Social Science	280	5.13%	1,003	2.93%
Earth Science	193	3.54%	680	1.99%
Other	8	0.15%	56	0.16%
Language and Literature	7	0.13%	34	0.10%

Document and QA shares differ because historical phases generated different numbers of questions per PDF. Domain labels describe the source PDF, not a separate topic inferred for each question.

Answer types and per-document sampling

Answer format	QA	Share
Integer	5,175	15.13%
Float	1,807	5.28%
List	11,500	33.61%
String	12,387	36.21%
Unanswerable	3,343	9.77%

There are 30,869 answerable and 3,343 unanswerable QA examples. The observed necessary-evidence page counts are: 0 pages: 3,343, 1 pages: 15,589, 2 pages: 10,220, 3 pages: 4,034, 4 pages: 1,022, 5 pages: 2, 6 pages: 1, 7 pages: 1.

Observed QA count per source document

QA per document	Documents
1	414
2	2,637
6	1
8	1
9	7
10	51
11	259
12	2,084
20	4

The 20-QA documents belong to the initial pilot; the two-QA limit applies to the current GPT-5.6 phase.

Generation methods

Recorded method	QA
legacy_full_document	32,432
released_prompt_window_v1	1,103
document_tool_agent_v1	677

Domain-label provenance

Recorded method	Documents
majority of full-context generator domain labels	4,330
arXiv primary category mapping	113
manual primary-topic classification for validated pilot batches	9
majority of independent full-document verifier topic labels	982
Crossref subject-query sampling; provisional until generator topic vote	24

Language annotation

Estimated document language	Documents	Document share	Associated QA
English (en)	5,453	99.908%	34,172
Spanish (es)	2	0.037%	14
French (fr)	1	0.018%	2
Russian (ru)	1	0.018%	12
Chinese (zh)	1	0.018%	12

Language labels are automatic estimates of the source document's dominant sampled language, not human-verified labels. We apply langid==1.1.6 to embedded PDF text from up to six pages spanning the document body. Non-English or uncertain predictions are rechecked after Unicode normalization and filtering short equation-like lines. Low-scoring or text-poor cases use und (undetermined). No OCR is run for this annotation. The normalized classifier score is not a calibrated accuracy estimate. QA prompts and generated answers are in English.

Per-document sampling pages, status, method, and score are available in metadata/document_languages.jsonl. The raw initial prediction is retained for rechecked cases.

Classification and annotation schema

Every admitted document has a primary-domain label. This single-label topical taxonomy is separate from the QA task taxonomy. Labels come from recorded generator votes, independent full-document verifier votes, arXiv primary-category mapping, provisional journal subject queries, or pilot assignments; counts and recorded methods appear in the provenance table above. These labels have not undergone comprehensive human validation.

The documents viewer and metadata/document_labels.parquet provide one row per source PDF: title, URL, source hash, document type, domain and its assignment method, automatic language annotation, page counts, unique image-token estimate, QA count, training-token sum, question types, generation models, and source license.

The default viewer provides one row per QA:

Field(s)	Meaning
id, document_id	Stable QA identifier and source-PDF join key
question, answer, answer_format	Question, rendered answer, and answer type; list answers are JSON strings in the viewer table
kind	single, multi_local, multi_distributed, reasoning, or unanswerable
answerable	Whether the complete source document supports an answer
source_domain, domain_tag_method	Primary topical domain and recorded assignment procedure
model, generation_phase	Actual recorded generator identifier and historical sampling phase
document_type, document_language	Source family and automatic language estimate; language provenance is in the document table
document_page_count, content_page_count	Complete PDF pages and content-bearing pages used by the production policy
image_paths	Ordered relative paths to every page image in the complete source PDF
token_count	Stored Qwen2.5-VL estimate for the complete serialized training sample
evidence_pages, evidence_page_count	Necessary evidence pages, using zero-based original PDF indices
evidence_span_normalized, evidence_distance_bin	Actual evidence span; separate from local/distributed task labels
related_context_pages	Inspection anchors for unanswerable questions, not answer evidence
source_url, source_license	Source attribution and article-level license

Question types

• single: factual extraction requiring one evidence page.
• multi_local: aggregate distinct necessary facts from multiple pages within an assigned contiguous 8-15-page window.
• multi_distributed: combine necessary evidence from separated positions in the full PDF; medium/far span constraints are recorded.
• reasoning: calculate, compare, count, or otherwise reason over grounded document evidence.
• unanswerable: the complete document does not supply the answer or a required operand; the exact target is Not answerable. and evidence_pages is empty.

For nonempty evidence, normalized span is (last_evidence_page - first_evidence_page) / (document_page_count - 1). Near is at most 0.2, medium is greater than 0.2 and at most 0.5, and far is greater than 0.5. Empty evidence is null / no_evidence, rather than a near-span example. A local sampling window can produce a medium normalized span in a short PDF; the sampling task and actual evidence distance should not be conflated.

Generation provenance

Recorded generation model	Documents	QA	QA share	B training tokens
gpt-5.5-2026-04-24	2,407	28,524	83.37%	1.745
gpt-5.6-sol	3,051	5,688	16.63%	0.334

Phase	Documents	QA	Maximum QA per document
Initial pilot (GPT-5.5)	4	80	20
GPT-5.5 production	2,403	28,444	12
Current GPT-5.6 production	3,051	5,688	2

The current snapshot is a mixed-model historical release, not an all-GPT-5.6 corpus. Use the model field when a model-specific subset is required.

The current production configuration uses GPT-5.6-sol, at most two QA examples per document, and one target QA at a time. New hybrid candidates pass structural checks and two independent full-document verification calls. The GPT-5.6 phase additionally applies a semantic-redundancy comparison across candidate questions from the same document. Historical examples retain their original recorded model and sampling policy.

The current hybrid recipe reuses the released window sampler, generation prompt, and two LongDocURL few-shot exemplars for single/local extraction and locally planned numeric reasoning. Distributed extraction, multi-page-planned numeric reasoning, and unanswerable tasks use a document-scoped GPT-5.6 API tool loop: overview, full-text search, page-text/image inspection, and bounded arithmetic. The CLI run was a workflow prototype; production records are generated through the GPT-5.6 API.

Tool agents have access to the complete document, but their recorded image views may be selective. Unanswerable agents must receive every extracted text page and inspect related page images before submission; two separate verification calls then receive all PDF page images. A failed literal search is never accepted as an absence proof. generation_method, generation_text_pages, teacher_pages, generation_model_turns, generation_tool_calls, and generation_trace record the actual process. verification_policy distinguishes the new two-check policy from historical records. Generation traces are archived with admitted examples. The final training context remains the complete ordered PDF for every task.

When running the included production code with journal_prefetch: true, start python code/data_synthesis_doc_vqa/journal_intake.py --root <corpus-root> --policy <policy.json> in a separate process alongside local_corpus.py run with the same root and policy. The supervised production controller starts and monitors intake automatically. The prepared-input buffer is bounded by both document count and disk space; the model worker remains the sole writer of the admitted-sample ledger.

Independent document prefetch downloads and renders complete journal PDFs while the model worker processes prepared documents. Generation, verification, retries, and semantic-redundancy review advance per document. Intake favors underrepresented domains by unique-PDF counts; this is a sampling policy, not a claim that the released corpus is already balanced. Local generation uses the released 8–15-page window; independent verification and the final training sample retain every PDF page.

Configured token-mixture weights are 36% single-page extraction, 18% local multi-page extraction, 18% distributed multi-page extraction, 18% reasoning, and 10% unanswerable. These are configuration weights, not fixed row quotas. Among answerable examples, the intended extraction/reasoning proportions correspond to 4:4:2. Observed counts and token totals are reported above.

The public pipeline and policy are included under code/. The original research code and training recipes are maintained on GitHub. The original paper's Seed 2.0/OCR pipeline and this public GPT-based pipeline are different releases.

Download and use

Load the QA or document table

from datasets import load_dataset

repo = "ZhaoweiWang/MMProLong-Data-Public"
qa = load_dataset(repo, "default", split="train")
documents = load_dataset(repo, "documents", split="train")

# Select the examples actually generated with GPT-5.6.
gpt56 = qa.filter(lambda row: row["model"] == "gpt-5.6-sol")

# Example topical/task filter.
physics_reasoning = qa.filter(
    lambda row: row["source_domain"] == "Physics" and row["kind"] == "reasoning"
)

The documents configuration is metadata with one row per PDF, not a second QA training split. Use document_id to join the two tables. Pin the dataset revision to a commit hash when reporting reproducible experiments.

Restore page images with bounded temporary disk usage

Use Python 3.12+ for the restoration scripts. Install the Hub CLI and table dependencies:

python -m pip install -U huggingface_hub datasets pyarrow

hf download ZhaoweiWang/MMProLong-Data-Public \
  --repo-type dataset --local-dir MMProLong-Data-Public \
  --include 'README.md' 'data/*' 'training/*' 'metadata/*' 'tools/*' \
            'corpus.jsonl.gz' 'release_manifest.json'

python MMProLong-Data-Public/tools/hydrate_assets.py \
  --root MMProLong-Data-Public --revision main --groups images

The helper downloads and verifies one archive at a time, extracts it, and removes that temporary archive. It refuses to overwrite a different existing file. Use --groups images sources provenance when PDFs and provenance records are also needed. Replace main with a fixed commit for reproducible restoration. Existing production ledgers are preserved.

Images are stored once per document. The QA table's image_paths and the training table's images are relative path strings, not embedded image bytes. Resolve them against the restored corpus directory on demand:

from pathlib import Path

root = Path("MMProLong-Data-Public")
page_paths = [root / path for path in qa[0]["image_paths"]]
assert all(path.is_file() for path in page_paths)

The canonical training/data.parquet stores VeOmni/LLaVA-style conversations with shared-image references. Adapt the image-loading step if a trainer expects bytes rather than paths. GPU training ingestion of this public snapshot has not yet been validated.

Download the complete backup

hf download ZhaoweiWang/MMProLong-Data-Public \
  --repo-type dataset --local-dir MMProLong-Data-Public
python MMProLong-Data-Public/tools/restore.py --root MMProLong-Data-Public

Check the current file sizes in release_manifest.json before downloading; full extraction also needs space for the unpacked assets. Every base and dated incremental archive listed in the current manifest is required. --verify-only checks the downloaded package without extracting it. In a fresh directory, add --resume to restore the SQLite ledger and discovery checkpoint from backup/. This backup contains admitted QA; incomplete production candidates and all historical raw model responses are not part of the release.

Files and formats

Path	Contents
data/train-*.parquet	QA viewer table with filtering labels and derived evidence fields
training/data.parquet	Canonical conversations and relative shared-image paths
corpus.jsonl.gz	Full admitted records, original answer types, evidence, verification records, and provenance
metadata/documents.jsonl	Original source-document metadata, authors, URLs, licenses, and PDF hashes
metadata/document_labels.parquet, .jsonl	One row per PDF with domain, language, document type, page counts, and associated QA statistics
metadata/document_languages.jsonl	Automatic language-detection evidence, status, and method
metadata/statistics.json	Exact totals, per-task statistics, quantiles, domain/model distributions, and histogram bins
metadata/asset_members.jsonl.gz	Individual archive-member paths, sizes, and SHA-256 hashes
assets/*.png	Public-corpus statistical figures used in this card
archives/images-*.tar	Ordered page images under images/<document_id>/pageNNN.png
archives/sources-*.tar	Source PDFs used by admitted QA
archives/provenance-*.tar	Source/license evidence and discovery metadata
backup/, code/	Production ledger/checkpoint and the public synthesis pipeline snapshot
package_manifest.json, release_manifest.json, checksums.sha256	Integrity records for the complete release
tools/	Checksum verification, restoration, and statistics utilities

Quality checks and limitations

• Every admitted QA passed structural validation and stored full-document model review. The exact PDF hash and all reviewed page indices are retained. GPT-5.6 document-level novelty checks are recorded where required.
• All referenced page images and source-PDF hashes are verified when each snapshot is prepared. Model reviews and automatic topic/language labels can still contain errors; they are not human gold annotations.
• This snapshot contains public research papers; the measured source and topic distributions above describe its coverage. New intake prioritizes underrepresented domains and supports PLOS and eLife journal articles alongside arXiv candidates. Books, manuals, and other document families from the original paper's pool are not represented by this release.
• Source PDFs were selected for article-level CC BY 4.0, with an explicit third-party-rights notice filter. Article-level licensing and text scanning are not exhaustive clearance of every figure. Original authors, URLs, licenses, and exceptions remain applicable. No new blanket license for synthetic QA is assigned by this intermediate release; the project's code license does not replace source-document rights.
• The held document 2609.29669v1 and its 20 pilot examples are excluded.
• Benchmark decontamination, comprehensive human review, and GPU training-ingestion checks remain incomplete for this public corpus. Do not transfer the original paper's completed checks or reported benchmark results to this separate dataset.

Reproduce the statistics

The profiling and annotation scripts are available on GitHub and mirrored in tools/.

python -m pip install numpy pyarrow matplotlib
python MMProLong-Data-Public/tools/profile_public_dataset.py \
  --snapshot-root MMProLong-Data-Public \
  --language-labels MMProLong-Data-Public/metadata/document_languages.jsonl \
  --output-dir public-corpus-profile

This reproduces aggregate statistics, labels, and figures from the small metadata/QA files. To recompute language annotations from source PDFs, install pymupdf requests langid==1.1.6 and run annotate_public_document_languages.py, followed by recheck_public_document_languages.py; both expose --help. The first pass streams verified tar archives in memory, and the recheck fetches only selected PDF byte ranges. Neither retains the PDF corpus on disk.

Statistics and annotations are computed from the admitted records in each published release. Exact token counts, source hashes, and file checksums are retained with the snapshot.

Contact

Open an issue in the MMProLong GitHub repository, or contact Zhaowei Wang at zwanggy@cse.ust.hk.

Citation

If you use this public dataset, its annotations, or the synthesis methodology, please cite MMProLong and identify this separately constructed public release and its revision in your experiment description.

@inproceedings{wang2026mmprolong,
  title={Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context},
  author={Wang, Zhaowei and Luo, Lishu and Duan, Haodong and Liu, Weiwei and Wu, Sijin and Luo, Ji and Yan, Shen and Peng, Shuai and Yuan, Sihang and Huang, Chaoyi and Lin, Yi and Song, Yangqiu},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS 2026)},
  year={2026},
  url={https://arxiv.org/abs/2605.13831}
}