Add polished dataset card with sampling + full dataset link

README.md

@@ -0,0 +1,96 @@
+---
+pretty_name: IPO Tables (HTML) — Random Yearly Sample
+tags:
+  - sec
+  - ipo
+  - s-1
+  - tables
+  - html
+  - finance
+  - information-extraction
+task_categories:
+  - text-classification
+  - other
+language:
+  - en
+license: mit
+size_categories:
+  - 1K<n<10K
+---
+
+# IPO Tables (HTML) — Random Sample Card
+
+A curated table extraction dataset from SEC filing documents, with raw table HTML plus provenance metadata.
+
+## What This Dataset Is
+
+This is a **random sample targeting 100 extracted tables per year** from filings in **1994–2026**.
+
+- Middle years are densely represented at 100 tables/year.
+- Edge years can be lower where fewer valid tables were available.
+- Tables are extracted directly from filing source files and stored as raw HTML.
+
+## Full Dataset Link
+
+The full dataset is available here:
+
+**https://drive.google.com/drive/folders/16hFhoXy52o9C2dARpxmBHQ6QCZjvum6S?usp=sharing**
+
+## Snapshot Stats
+
+- Year range: 1994–2026 (33 years)
+- Total extracted tables: 3,099
+- Years with 100 tables: 30
+- Underfilled edge years:
+  - 1994: 35
+  - 1995: 64
+  - 2026: 0
+
+## Files
+
+- `sampled_table_html.csv`
+- `sampled_table_html.parquet`
+- `year_summary.csv`
+- `dataset_info.json`
+
+## Row Schema
+
+Each row contains:
+
+- `sample_id`: unique row id
+- `target_year`: sampled year bucket
+- `source_file`: source filing path used for extraction
+- `source_filename`: filing filename
+- `source_extension`: file extension (`.htm`, `.html`, `.txt`)
+- `cik`: parsed company identifier (when available)
+- `accession_from_filename`: accession-like id parsed from filename
+- `table_pick_index_in_filing`: selected table index in filing
+- `table_start_char`: char offset start in source text
+- `table_end_char`: char offset end in source text
+- `table_html`: raw extracted table HTML
+- `table_html_char_length`: HTML character length
+
+## How It Was Built
+
+1. Enumerate filing files by year from local corpus.
+2. Shuffle filings with deterministic seed.
+3. Extract `<table>...</table>` blocks using HTML regex matching.
+4. Keep up to one table per filing for diversity.
+5. Continue until 100 tables for the year (or exhaust candidates).
+
+## Intended Use
+
+- Table parsing benchmarks
+- Financial document IE experiments
+- HTML-to-structured-table extraction
+- LLM/RAG evaluation on filing tables
+
+## Limitations
+
+- Extraction is HTML-pattern based; malformed markup can reduce recall.
+- Some provenance fields depend on filename conventions.
+- Edge years may contain fewer tables than the 100 target.
+
+## Citation
+
+If you use this sample card dataset, please cite the project/repository and include the full dataset link above.