bn_en_banglish_100k_finetune

A curated, category-balanced, deduplicated 100,000-row subset of tensorlabco/bn_en_banglish_v2 (1,370,553 source rows), built for Bangla↔English translation fine-tuning. This is the full, 3-text-field variant; see the companion tensorlabco/bn_en_100k_finetune for a bn_text/en_text-only projection of the exact same 100,000 rows and split.

Columns

Column	Description
category	Fine-grained topic (e.g. sports, bangladesh, book)
type	Coarse grouping: news or story (story maps 1:1 to category == "book")
bn_text	Bangla text
en_text	English text
banglish_text	Romanized Bangla / code-mixed "Banglish" text

Construction

1. Word-count filters. story rows keep bn_text word count ≥ 100 (no upper bound — story is naturally clustered around 500 words). news rows are filtered to bn_text word count in [300, 600]; any news category with fewer than 30 rows in that window is dropped entirely.
2. Pool-level deduplication and quality filtering, before any sampling decision. The filtered pool is deduplicated (exact-match, then MinHash/LSH near-duplicate collapse at Jaccard ≥ 0.9) and rows with a banglish_text length wildly disproportionate to bn_text are dropped — both done at the pool level so every category's available supply is a true count of unique, usable content before quotas are decided, and the published banglish_text column is never null.
3. Balanced category quotas via water-filling. story (book) takes its full cleaned supply; the remaining budget is distributed across the 19 news categories, smallest-supply categories getting their full available count and the rest capped at a shared ceiling so the total lands exactly on 100,000.
4. Stratified 80/10/10 split, independently per category, so the category mix is preserved in every split.

Verified on the final selection: 0 near-duplicate bn_text pairs (both within-split and cross-split), 0 train/validation/test leakage, 0 banglish_text outliers, 0 empty fields.

Dataset structure

100,000 rows total: 80,000 train / 10,000 validation / 10,000 test, across 20 categories.

category	train	validation	test	total
book	22,547	2,819	2,815	28,181
news	4,775	597	597	5,969
entertainment	4,775	597	597	5,969
sports	4,775	597	597	5,969
bangladesh	4,775	597	597	5,969
lifestyle	4,775	597	597	5,969
world	4,775	597	597	5,969
education	4,775	597	597	5,969
business	4,775	597	597	5,969
politics	4,254	532	532	5,318
opinion	4,251	531	532	5,314
old-prothomalo	4,206	526	525	5,257
technology	3,808	476	476	4,760
onnoalo	991	124	124	1,239
chakri	897	112	112	1,121
religion	307	38	39	384
fun	260	32	33	325
special-supplement	178	22	22	222
anniversary	75	9	10	94
durporobash	26	3	4	33

Known limitations

• book (28,181 rows) is short of the news categories' shared ceiling and short of its own pre-cleaning pool (29,328) — this is real: the source has ~1,147 duplicate or banglish-outlier story rows and no further unique/usable story content exists in the source to replace them with. The 8 largest news categories absorbed the corresponding extra budget instead.
• The smallest retained categories (durporobash: 33, anniversary: 94) have single-digit validation/test counts — kept anyway to preserve every sub-category with meaningful supply rather than dropping thin ones for cleaner statistics.

Source

Derived from tensorlabco/bn_en_banglish_v2 (all 3 original splits combined and treated as one pool before re-filtering and re-splitting).