Add dataset card

README.md

@@ -1,23 +1,22 @@
 ---
-dataset_info:
-  features:
-  - name: image
-    dtype: image
-  - name: target
-    dtype: string
-  - name: gt_xml
-    dtype: string
-  - name: record_id
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 44465018
-    num_examples: 428
-  - name: test
-    num_bytes: 5667193
-    num_examples: 58
-  download_size: 49835021
-  dataset_size: 50132211
+license: mit
+task_categories:
+- image-to-text
+language:
+- fr
+tags:
+- ocr
+- structured-extraction
+- kie
+- index-cards
+- glam
+- nuextract
+- handwriting-recognition
+size_categories:
+- n<1K
+source_datasets:
+- Teklia/DAI-CReTDHI-IndexCards-KIE
+pretty_name: Teklia Index-Card KIE (flat schema, train/test)
 configs:
 - config_name: default
   data_files:
@@ -26,3 +25,25 @@ configs:
   - split: test
     path: data/test-*
 ---
+
+# Teklia Index-Card KIE — flat schema (train / test)
+
+French handwritten **vital-records index cards** (Archives of Tours), prepared for schema-conditioned
+structured extraction. Repackages [`Teklia/DAI-CReTDHI-IndexCards-KIE`](https://huggingface.co/datasets/Teklia/DAI-CReTDHI-IndexCards-KIE)
+with a **flat-by-role JSON target** derived deterministically from the nested XML ground truth.
+
+| split | cards |
+|---|---|
+| `train` | 428 |
+| `test` | 58 |
+
+Births, marriages and deaths; children, spouses and parents, each routed to its role. Training and
+evaluation data for the [index-card-extractor-4b](#) research model.
+
+## Columns
+`image` (IIIF JPEG) · `target` (flat-by-role JSON — the extraction label) · `gt_xml` (nested XML
+ground truth) · `record_id`.
+
+## Provenance & license
+Derived from `Teklia/DAI-CReTDHI-IndexCards-KIE` (MIT); released under **MIT**. Please cite the
+original Teklia dataset.