Unichart Pretrain Data: 6.9 Million Chart Image-Text Pairs

6,898,333 rows of chart images paired with text queries and labels, hosted on Hugging Face by ahmed-masry and last updated in March 2024. The dataset is structured for training multimodal models, with each row containing an image name, an input query, and an output label. Its primary use appears to be pretraining models for chart understanding and generation tasks.

Use Cases

• Pretraining multimodal AI models based on the image-text-label structure.
• Training chart question-answering systems based on the query and label fields.
• Fine-tuning models for chart image captioning or description generation based on the query-label relationship.
• Benchmarking model performance on chart understanding tasks based on the provided labels.

Strengths

• Contains 6,898,333 rows, providing a substantial scale for pretraining.
• Features a defined structure with 'imgname', 'query', and 'label' columns, suggesting organized multimodal data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The source and methodology for data collection are not specified, which may affect reproducibility.

Provenance

Source

ahmed-masry on Hugging Face

Freshness

Last updated 2024-03-23 06:30:42.