trannhiem/TranNhiem-Vietnamese-DocumentImage-Reasoning

TranNhiem Vietnamese Document-Image Reasoning (V-Doc)

[image]

Vietnamese document-image understanding with explicit reasoning: multi-turn question–answering grounded on scanned/rendered Vietnamese document pages (textbooks, articles, worksheets). Each answer includes a step-by-step chain-of-thought. Reasoning and Answer was synthesized by Qwen3.5 over the Viet-Doc-VQA-II document collection.

• —Curated by: Trần Nhiệm..
• —Languages: Vietnamese (vi) answers · English (en) reasoning
• —Modality: document image + text → text
• —Records: 64,516 document-grounded multi-turn conversations
• —Images: 64,516 document page images
• —Image shards: 6 × ~5 GB tar files under shards/

🔎 Size & configs

⚠️ Full dataset = 64,516 records / 64,516 document images (~9.5 GB). The Dataset Viewer shows a 300-row `preview` for quick visual inspection — it is not the full size. Total download size is shown on the Files tab.

• —`preview` (default, shown in the viewer) — 300-row sample with rendered image thumbnails + flat columns (image, question, model_reasoning, model_answer).
• —`full` — all 64,516 records (id, image path, width, height, n_pairs, conversations as JSON). Pick it from the config dropdown, or load_dataset("trannhiem/TranNhiem-Vietnamese-DocumentImage-Reasoning", "full").
• —Images: shards/images-*.tar (extract → images/..., matching the image field).

👀 Preview config columns

The `preview` config is a 300-row sample flattened into clean columns:

Switch to the `full` config for the complete nested multi-turn conversations.

Dataset Summary

Each record is a document page image plus a multi-turn dialogue (typically 5 Q&A pairs, 10 turns). Questions require reading and reasoning over the document content — reading comprehension of Vietnamese text, extracting facts/lists from passages, interpreting tables and worksheet exercises, and solving embedded problems (e.g., arithmetic in textbook pages). Every model turn contains both a reasoning trace and a final answer, targeting document-grounded reasoning rather than shallow OCR.

📊 Dataset Statistics

Token counts estimated with the Qwen3 tokenizer over a 20,000-record sample; totals extrapolated to all 64,516 records.

Scale

• —Records (document pages): 64,516
• —Q&A pairs: ~323K (avg 5.0 pairs/record, 10 turns)
• —Estimated total text: ~0.24 billion tokens — reasoning ≈ 0.16 B · answers ≈ 0.08 B · questions ≈ 0.01 B
• —Avg tokens/record (all turns): ~3,790

Per-field length — tokens (mean / median / p90)

Document questions run longer than natural-image ones (they quote passage text), and reasoning is the largest field — consistent with reading-comprehension over dense document pages.

Images (document pages)

• —Total size: 9.5 GB ·
• —Dimensions: mean 942 × 1268 px (median 794 × 1123), up to 1620 × 2184
• —Orientation: 93% portrait · 7% landscape (typical A4-style document pages)

Supported Tasks

• —Document Visual Question Answering (DocVQA) — Vietnamese, open-ended.
• —Document-grounded chain-of-thought reasoning (reading comprehension, table/worksheet reasoning, math).
• —Multi-turn document dialogue.
• —Instruction tuning for Vietnamese document-understanding VLMs.

Languages

Questions and answers are Vietnamese; the reasoning field is English (generator working language).

Dataset Structure

Fields (one JSON object per line, in data/vdoc.jsonl)

Each gpt turn:

{"from": "gpt", "value": {"reasoning": "<English CoT>", "answer": "<Vietnamese answer, Markdown>"}}

Files & layout

preview/preview-000.parquet   # 300-row viewer sample (image + question + reasoning + answer)
data/vdoc.jsonl               # 64,516 records
shards/images-00000.tar ...   # 6 shards; members images/<original-relative-path>.jpg
README.md

Usage

huggingface-cli download trannhiem/TranNhiem-Vietnamese-DocumentImage-Reasoning \
    --repo-type dataset --local-dir ./vdoc
cd vdoc && for t in shards/images-*.tar; do tar -xf "$t"; done   # -> images/...

import json; from PIL import Image
rec = json.loads(open("vdoc/data/vdoc.jsonl").readline())
img = Image.open(f"vdoc/{rec['image']}")
for t in rec["conversations"]:
    if t["from"] == "gpt":
        print(t["value"]["reasoning"], "\n", t["value"]["answer"])

Dataset Creation

• —Source documents: Viet-Doc-VQA-II — Vietnamese document page images (textbooks/articles/worksheets).
• —Generation: questions + multi-turn reasoning/answers synthesized with Qwen3.5.
• —Packaging: images stored as WebDataset-style tar shards; image paths rewritten to images/<relpath>.

Considerations & Limitations

• —Model-generated: reasoning/answers may contain factual/OCR/arithmetic errors — treat as silver supervision and verify before high-stakes use.
• —Bilingual asymmetry: reasoning EN, answers VI (by design).
• —Source/rights: document images derive from the Viet-Doc-VQA-II collection; users must comply with the underlying source terms.
• —Bias: inherits the source document distribution + Qwen3.5 generator biases.

License

Research use (license: other); underlying documents follow their original source terms.

Citation

@misc{trannhiem_vdoc_reasoning,
  title  = {TranNhiem Vietnamese Document-Image Reasoning (V-Doc)},
  author = {Tran, Nhiem},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/trannhiem/TranNhiem-Vietnamese-DocumentImage-Reasoning}}
}