---
license: other
task_categories:
  - translation
  - image-to-text
language:
  - bs
  - en
pretty_name: Lilly evaluation and training bytes
---

# Lilly data (the bytes git does not hold)

Code, answer keys, credits and results stay in
[ssaaffaakk/Lilly](https://github.com/ssaaffaakk/Lilly). This dataset is the
photographs and the cleaned parallel sentences those files name.

| File | What it is |
|---|---|
| `translation-clean.zip` | `train.tsv` (313,612 pairs), `valid.tsv`, `test.tsv`, `ntrex-holdout.tsv` — WikiMatrix / SETIMES / TED2020 / Tatoeba after the project's filters. The base model already saw this material (`docs/WHITE-PAPER.md` §3.1). |
| `ocr-the-40.zip` | The 40 Wikimedia Commons photographs the product OCR number (67.0%) was measured on, plus `commons40-sha1.json`. |
| `ocr-test-v2.zip` | test-v2: 280 Commons photographs (132 with text), credits, the two blind passes, the answer key. |
| `ocr-harvest.zip` | Harvested Commons sign photographs and `CREDITS.tsv`. |
| `ocr-crops.zip` | Crops cut from the harvest, with `labels-human.tsv`. |

Photograph licences are per-file in each archive's `CREDITS.tsv` (Wikimedia Commons).
The parallel sentences come from OPUS (SETIMES, WikiMatrix, TED2020, Tatoeba) and
NTREX-128; credit those sources, not this page.

Not here: FLORES-200 (fetch from Meta), FLEURS (fetch from Google), the 20,240
Mapillary frames (`afaksrmeli/lilly-mapillary-photos` on Kaggle — not a train set).