BDRC/script-classification-benchmark-balance-subclass

Script Classification Benchmark (Balanced Subclasses)

Holdout benchmark for six-class Tibetan script classification (360 page images). Each class contains 60 real BDRC manuscript scans from Data/benchmark/.

Class Distribution

Class	Images
Danyig	60
Druma	60
Gyuyig	60
Pedri	60
Tsugdri	60
Uchen	60

Subclass Breakdown & Source Pool Availability

Every subclass is perfectly balanced and contains an equal portion of its parent class's images. For smaller classes with limited distinct works (such as Gongshabma and Uchen SugRing), the benchmark employs a work-reuse fallback to select additional distinct pages from the same manuscripts, ensuring subclass balance is perfectly equal without under-filling.

Parent Class	Subclass	Benchmark Count	Total Source Pool
Danyig	DraDring	12	309
	DraRing	12	277
	Drathung	12	1,632
	Gongshabma	12	94
	Tsegdrig	12	3,161
Druma	Dhumri	12	471
	DruDring	12	735
	DruRing	12	569
	Druchen	12	97
	Druthung	12	1,067
Gyuyig	Khyuyig	20	675
	Tsumachug	20	900
	Yigchung	20	768
Pedri	Peri	30	2,545
	Petsuk	30	6,206
Tsugdri	Trinyig	20	301
	Tsugchung	20	525
	Tsugthung	20	93
Uchen	Uchen SugDring	20	5,013
	Uchen SugRing	20	79
	Uchen Sugthung	20	793

Parquet schema

Column	Type	Description
id	string	BDRC page id
image_bytes	binary	JPEG/PNG page image
script_type	string	One of the six script families (parent)
script_name	string	Hierarchical sub-script name (e.g. Uchen Sugthung)
source	string	bdrc (real scans only)
font_name	string	Empty for BDRC scans

Load in Python

from datasets import load_dataset

repo = "BDRC/script-classification-benchmark-balance-subclass"
ds = load_dataset(repo, split="benchmark")
print(len(ds), ds.column_names)  # 360, ['id', 'image_bytes', 'script_type', 'script_name', 'source', 'font_name']

row = ds[0]
img = row["image_bytes"]  # bytes — same schema as BDRC training Parquet
print(row["source"], row["font_name"])

Page-level BDRC holdout ids for training exclusion live in benchmark_page_ids.json in the local Data/benchmark/ folder.