Chart2Code Quality Attention

Training data used for the yifei-12/charttocode model. It contains 11,712 chart-to-Matplotlib examples together with the chart images and the auxiliary attention-supervision artifacts used by positive-teacher KL training.

Files

• chart2code_quality_attn.json: ShareGPT-style user/assistant examples.
• archives/images_sharded.tar: the 11,712 chart images referenced by the dataset JSON, stored as images_sharded/<prefix>/<id>.png after extraction.
• visual_token_masks.json: visual-token supervision masks.
• quality_block_index.json: chart-element block annotations.
• archives/enhanced_rois_sharded.tar: title/legend ROI crops, plus enhanced_rois/enhanced_roi_index.json after extraction.

Extract both archives at the repository root before training:

tar -xf archives/images_sharded.tar
tar -xf archives/enhanced_rois_sharded.tar

All image and ROI references then resolve as repository-relative paths. Image IDs are sharded by their first three digits to stay within Hugging Face's per-directory file limit.

Load the records

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="chart2code_quality_attn.json",
    split="train",
)

The records use messages and images fields. Resolve entries in images relative to the downloaded repository root.

Related model

The corresponding fine-tuned model is available at yifei-12/charttocode.