trannhiem/TranNhiem-Vietnamese-ImageText-Reasoning

TranNhiem Vietnamese Image-Text Reasoning (V-LAION)

[image]

Large-scale Vietnamese multimodal reasoning: multi-turn visual question–answering grounded on natural images, where every answer ships with an explicit chain-of-thought. Reasoning traces and Answer were synthesized by Qwen3.5 over images from the LAION-derived Vi-Laion-gemini-VQA set.

• —Curated by: Trần Nhiệm
• —Languages: Vietnamese (vi) answers · English (en) reasoning
• —Modality: image + text → text
• —Records: 544,795 image-grounded multi-turn conversations
• —Images: 544,795 (PNG, a few JPG)
• —Image shards: 50 × ~5 GB tar files under shards/

🔎 Size & configs

⚠️ Full dataset = 544,795 records / 544,795 images (~268 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 544,795 records (id, image path, width, height, n_pairs, conversations as JSON). Pick it from the config dropdown, or load_dataset("trannhiem/TranNhiem-Vietnamese-ImageText-Reasoning", "full").
• —Images: shards/images-*.tar (extract → images/<id>.png, matching the image field).

👀 Preview config columns

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

Switch to the `full` config to browse the complete annotations (nested multi-turn conversations).

Dataset Summary

Each record is one image plus a multi-turn dialogue (typically 5 Q&A pairs, 10 turns) probing the image from several angles — object/material ID, on-image text (OCR), event/context recognition, colours and connotations, image quality/provenance, and secondary-object roles. Every model turn contains both a reasoning trace and a polished answer, targeting reasoning VLMs rather than short-answer accuracy alone.

📊 Dataset Statistics

Token counts estimated with the Qwen3 tokenizer over a 20,000-record sample; totals extrapolated to all 544,795 records.

Scale

• —Records (images): 544,795
• —Q&A pairs: ~2.72 million (avg 5.0 pairs/record, 10 turns)
• —Estimated total text: ~2.14 billion tokens — reasoning ≈ 1.15 B · answers ≈ 0.94 B · questions ≈ 0.05 B
• —Avg tokens/record (all turns): ~3,930

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

Images

• —Total size: 268 GB
• —Dimensions: mean 713 × 579 px (median 660 × 510), up to 1920 × 1824
• —Orientation: 71% landscape · 16% portrait · 13% square

Supported Tasks

• —Visual Question Answering — Vietnamese, open-ended, long-form.
• —Multimodal chain-of-thought reasoning — train/evaluate explicit visual reasoning.
• —Multi-turn visual dialogue — context across 5 Q&A pairs per image.
• —Instruction tuning for Vietnamese vision-language assistants (LLaVA-style).

Languages

Questions and answers are Vietnamese; the reasoning field is English (the generator's working language) — a deliberate bilingual design (reason in EN, answer in VI).

Dataset Structure

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

Each gpt turn is a structured object:

{"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/subset-1.jsonl           # 437,205 records
data/subset-2.jsonl           # 107,590 records
shards/images-00000.tar ...   # 50 shards, members images/<id>.png
README.md

Usage

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

import json; from PIL import Image
rec = json.loads(open("vlaion/data/subset-1.jsonl").readline())
img = Image.open(f"vlaion/{rec['image']}")
for t in rec["conversations"]:
    if t["from"] == "gpt":
        print(t["value"]["reasoning"], "\n", t["value"]["answer"])

Dataset Creation

• —Source images: Vi-Laion-gemini-VQA (Vietnamese-oriented VQA images derived from LAION).
• —Generation: questions + multi-turn reasoning/answers synthesized with Qwen3.5.
• —Packaging: images stored as WebDataset-style tar shards; image paths rewritten to images/<id>.png.

Considerations & Limitations

• —Model-generated: reasoning/answers may contain factual errors, hallucinations, or OCR mistakes — treat as silver supervision and spot-check.
• —Bilingual asymmetry: reasoning EN, answers VI (by design).
• —Source/PII: images derive from web (LAION) data and may show public figures/brands/text; users must comply with the underlying image rights.
• —Bias: inherits LAION distribution + Qwen3.5 generator biases.

License

Research use (license: other); underlying images follow their original LAION/source terms.

Citation

@misc{trannhiem_vlaion_reasoning,
  title  = {TranNhiem Vietnamese Image-Text Reasoning (V-LAION)},
  author = {Tran, Nhiem},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/trannhiem/TranNhiem-Vietnamese-ImageText-Reasoning}}
}