varadsrivastava/financial-chart-ocr

Financial Chart OCR

Financial charts paired with gold markdown tables, for chart-to-table extraction and for deriving chart question-answering supervision. The release accompanies the ICAIF '26 paper From Curriculum to Rehearsal Learning: Towards Small Financial Chart VLMs, and includes the paper's 50-chart held-out extraction benchmark.

config	split	charts	what it is
natural	train	199	real charts cropped from public financial disclosures
natural	benchmark	50	real charts held out from every model trained in the paper
synthetic	train	2,677	charts rendered with Matplotlib from generated tables
	total	2,926

from datasets import load_dataset
bench = load_dataset("varadsrivastava/financial-chart-ocr", "natural", split="benchmark")
syn   = load_dataset("varadsrivastava/financial-chart-ocr", "synthetic", split="train")

Fields

field	description
image	the chart image
markdown	gold annotation: title, any sub-headings and units, the data table(s), and surrounding context where the source provides it
chart_id	UUID (real charts) or generation timestamp (synthetic)
source_type	named (source issuer recorded), screenshot (source not recorded), or synthetic
issuer	normalised issuer slug for named charts (e.g. goldman, merck), otherwise null
source_tag	the capture name the chart was stored under

Real charts (natural)

Sources. Crops from publicly available corporate disclosures: annual reports, ESG and sustainability reports, investor presentations and broker research, across banking, technology, consumer, industrials, insurance, healthcare and energy issuers in several regions. Images are at roughly native resolution (median 766×538 px).

Annotation. Each chart was transcribed into markdown by Claude Opus 4.6 and then checked side by side against the image by one of two annotators in a custom annotation tool, who corrected titles, headers, units and values where needed. The set was split between the two annotators, so each chart was verified once; corrections were not logged, and no inter-annotator agreement is available. Treat the tables as human-verified, not as double-annotated.

Benchmark composition (50 charts, 38 issuers), by primary chart type: 13 bar–line combinations, 12 line, 8 stacked bar, 8 pie or donut, 7 simple or grouped bar, 2 multi-panel bar. The benchmark is disjoint from train by chart but not by issuer: 26 of its charts come from an issuer that also has charts in train. A 256-bit perceptual-hash comparison finds no near-duplicates between the two splits.

Synthetic charts (synthetic)

Each chart belongs to one of ten categories picked at random: financial, geospatial, technology, retail, manufacturing, energy, transportation, entertainment, agriculture and real estate. For each, Claude Opus proposed column schemas (names, units, value types); random number generators then drew the values, with 5 to 25 data points per chart. Claude Sonnet 4.6 wrote the Matplotlib 3.10.3 code that renders each chart, over thirteen predefined chart types: bar, horizontal bar, grouped bar, stacked bar, bar–line combination, line, area, pie, donut, waterfall, scatter/bubble, heatmap and histogram. Because the values are generated, each table is the exact ground truth for its chart.

The type mix was not logged. In a random sample of 48 renders: bar–line combinations (usually dual-axis) ≈40%, simple and grouped bars ≈20%, stacked bars ≈15%, line ≈13%, and the rest other types. Each annotation holds the generating table followed by descriptive sections ("Dataset Overview", "Key Insights", …).

Renders are large (median ≈8.6 MP against 0.41 MP, about 20× a typical real chart) and print their data labels. Under a fixed vision-token budget they are downscaled far more than real charts; see the paper for the consequences. Synthetic tables are longer than real ones (median 13 rows against 6) with a heavy tail; the paper caps training targets at 1,024 tokens.

Intended use

Research on chart-to-table extraction and chart question answering, in particular as a source of tables from which question–answer supervision can be derived by program. The 50-chart benchmark split should not be used for training if results on it are to be compared with the paper.

Evaluation harness

The eval/ folder holds the code that produced the paper's FinChart-Bench and extraction numbers: the prompts, greedy decoding settings, answer extractor and extraction scorer, with reference scores. Run python -m finchart_eval.extraction --adapter varadsrivastava/FinChart-VL-4B to score the released adapter on the natural benchmark split.

Licensing and copyright

Copyright in the real chart images remains with their original publishers. They are redistributed for non-commercial research use only, following the practice of comparable benchmarks built from public financial filings. If you are a rights holder and want a chart removed, open a discussion on this repository and it will be taken down.

The annotations and the synthetic charts are released under CC BY-NC 4.0.

The charts report aggregate corporate and market figures; the dataset contains no personal data.

Limitations

• 249 real charts is small; the 50-chart benchmark supports paired comparisons better than absolute claims.
• Each real-chart table was verified by a single annotator after an LLM draft.
• English-language charts only; sectors that publish detailed quantitative disclosures are over-represented.
• The synthetic chart-type shares are estimates from a 48-chart sample, not from generator logs.

Related

• Model: varadsrivastava/FinChart-VL-4B (the model reported in the paper)

Citation

@inproceedings{srivastava2026finchartvl,
  title     = {From Curriculum to Rehearsal Learning: Towards Small Financial Chart {VLMs}},
  author    = {Srivastava, Varad and Shukla, Abhay and Casas Hernandez, Victor},
  booktitle = {Proceedings of the ACM International Conference on AI in Finance (ICAIF '26)},
  year      = {2026}
}

Disclaimer

This work was carried out at AI Lab, Barclays. The views expressed here are those of the authors and do not represent the views of Barclays.