lducc/vlsp2026-chart-cards

Chart cards for VLSP 2026 Numerical Reasoning QA

Chart cards used by Team Arcane's system in the VLSP 2026 shared task on multimodal numerical reasoning over Vietnamese financial reports. The system ranked second on the private leaderboard (78.66 Program Accuracy). The submitted adapter is lducc/vlsp2026-arcane-writer.

• Code: github.com/lducc/vlsp
• Paper: Retrieval-Augmented Fine-Tuning for Vietnamese Multimodal Numerical Reasoning
• Task page: VLSP 2026 Numerical Reasoning QA

Chart cards

A chart card lists the title, series labels and category labels of one chart image. Programs in this task select chart values by label, as in chart_at(Image 1; VN-Index; Apr-20; none), so the writer model needs the labels and reads these cards in place of the images. The cards are the only model-generated data in the system and appear in both training and test prompts.

Example: Image 1 of a training report and its card.

{"title": "Dữ liệu quan trọng", "series_labels": ["VHC VN", "VN-Index"], "x_axis_labels": ["Apr-20", "Jul-20", "Oct-20", "Jan-21"]}

Files

Split	File	Charts
train	train_cards.jsonl	187
public_test	public_cards.jsonl	61
private_test	private_cards.jsonl	83

charts.py is the script that generated the cards and contains the full reader prompt.

Each line is one chart:

Field	Content
image	File name of the chart in the official task data
raw	Text returned by the chart reader
parsed	The card as JSON with title, series_labels and x_axis_labels

Generation

Setting	Value
Model	Qwen/Qwen3-VL-8B-Instruct (8B parameters), revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
Image input	Chart enlarged to twice its stored size with Lanczos resampling
Decoding	Greedy (temperature 0), at most 1,024 new tokens, one pass per chart
Inference	vLLM, bfloat16

The reader is asked to copy visible text exactly, including Vietnamese diacritics and date formats, and to leave out axis tick values. The chart images belong to the official task data and are not included.

Usage

from datasets import load_dataset

cards = load_dataset("lducc/vlsp2026-chart-cards", split="train")
print(cards[0]["image"], cards[0]["parsed"])

Citation

@misc{arcane2026chartcards,
  title  = {Chart cards for the {VLSP} 2026 Numerical Reasoning {QA} shared task},
  author = {Le, Anh Duc and Pham, Duc Cong and Nguyen, Huu Thanh},
  year   = {2026},
  url    = {https://huggingface.co/datasets/lducc/vlsp2026-chart-cards}
}

@misc{arcane2026vlsp,
  title  = {Retrieval-Augmented Fine-Tuning for Vietnamese Multimodal Numerical Reasoning},
  author = {Le, Anh Duc and Pham, Duc Cong and Nguyen, Huu Thanh},
  year   = {2026},
  note   = {System paper, VLSP 2026 shared task on Numerical Reasoning QA},
  url    = {https://github.com/lducc/vlsp}
}