README.md

---
pretty_name: "EIS-Text250: 1970s U.S. Environmental Impact Statements (text-only, 250 docs)"
language:
  - en
license: other
license_name: mixed-see-readme
task_categories:
  - text-generation
  - document-question-answering
  - summarization
tags:
  - document-understanding
  - historical-documents
  - ocr
  - environmental-impact-statement
  - nepa
  - government-documents
  - long-documents
size_categories:
  - 10K<n<100K
---

# EIS-Text250: 1970s U.S. Environmental Impact Statements (text-only)

Per-page OCR/extraction **text** for 250 scanned 1970s U.S. federal
**Environmental Impact Statements (EIS)** from the Northwestern University
Library collection — the text-only companion to
[`Windsao/eis-subset50`](https://huggingface.co/datasets/Windsao/eis-subset50)
(which carries full page images for a 50-doc subset). Built to test how current
models handle long, dense, historical government text: mean ~300 pages/doc,
1970s typewriter prose, OCR noise from degraded microfilm.

## Files

- `text.jsonl` — one row per page, all 250 documents:

```json
{
  "doc":       "p1274_35556039648936",
  "page":      12,
  "model":     "marker",            // extraction pipeline: "marker" or "chandra"
  "text":      "...page text...",
  "image_url": "https://nu-impulse-production.s3.amazonaws.com/..."
}
```

- `docs.txt` — 250 rows, tab-separated: `doc_id  models  pages  group`
  (`group` = `subset50` if the doc is also in the image subset, else `text-only`).

No images are included; `image_url` points to the publicly readable full-page
scan (~2400×3200 JPG) on S3 if you need pixels for a specific page. For 50 of
these documents, all page images are packaged in `Windsao/eis-subset50`.

## Selection

250 of the 277 documents with high-quality text extraction in the corpus
(marker or chandra pipelines): all 50 docs of `eis-subset50`, plus the 200
largest remaining docs by page count (the 27 shortest were dropped).

- `marker` pages: text flattened from marker's block tree (HTML stripped).
- `chandra` pages: raw markdown **including model-generated figure alt-text**.
- Pages covered by both pipelines use chandra.

## Known issues / gotchas

- Page 1 of many documents is a **microfilm resolution test chart**, not content.
- A small fraction of pages are legitimately empty (blank microfilm frames).
- OCR quality varies with scan degradation; text may contain OCR errors —
  treat as noisy ground truth.
- `marker` emits no figure descriptions; only `chandra` docs describe figures.

## Provenance & licensing

- **Source documents**: U.S. federal government EIS filings (1970s) — works of
  the U.S. federal government are in the **public domain** (17 U.S.C. § 105).
- **Scans**: Northwestern University Library digitization (Impulse pipeline).
- **OCR/extraction**: marker and chandra pipelines over the scans.
- Compiled by Shang Wu (Northwestern) for research on long-document
  understanding. Please cite this dataset page if you use it.