Annual financial reports are long, table-heavy, and use domain-specific Vietnamese language that most public datasets lack. This collection provides OCR-extracted full-texts from 18,231 annual reports (2015–2025) across 1,491 stock tickers, giving practitioners a large, domain-focused corpus for document understanding and extraction tasks.
What Sets It Apart
- Scale & coverage: 18,231 OCRed reports spanning banks, securities, insurance and corporate filings across 2015–2025 — useful for temporal studies and cross-company comparisons.
- OCR-first format: data is delivered as plain TXT OCR outputs preserving layout-derived text; the dataset reports ~95% accuracy on numeric and table regions (useful when training table-aware extractors or evaluating OCR post-processing).
- Hierarchical organization: files are organized by ticker → year → report type, which simplifies per-company retrieval and dataset slicing for time-series or per-sector experiments.
- Focused domain: unlike generic OCR corpora, content is exclusively Vietnamese financial statements, so models trained or validated here learn domain-specific formatting, terminology and table patterns.
Who It's For and Tradeoffs
Great fit if you want OCR benchmarking on Vietnamese financial docs, pretraining/fine-tuning VLMs for document QA, building RAG systems over company reports, or extracting financial indicators across many issuers. Look elsewhere if you need original scanned PDFs, guaranteed error-free structured tables, or a permissive commercial license (this dataset is CC BY‑NC 4.0). Also expect some OCR noise in poor-quality scans — downstream pipelines should include validation or correction steps.
