ChatgaiyyaBench

A multi-task benchmark and unified corpus for Chittagonian (চাটগাঁইয়া) — a Bangla dialect spoken by tens of millions of people with almost no NLP resources.

Pipeline, models and full results: https://github.com/A-K-M-Asifuzzaman/Chatgaiya-AI

The benchmark — 13,497 items, 8 tasks

Task	Items	Evidence	Dialectness low / mid / high
classification_vulgarity	3,015	silver	108 / 1,307 / 1,600
romanized	2,499	gold	124 / 1,214 / 1,161
translation	2,222	gold	102 / 968 / 1,152
semantic_similarity	2,109	gold	91 / 918 / 1,100
sentiment	1,767	silver	38 / 587 / 1,142
linguistic_diagnosis	708	silver	30 / 297 / 381
qa_cloze	590	silver	19 / 254 / 317
spelling_normalization	587	gold	19 / 249 / 319

7,417 gold · 6,080 silver. Read that column before using any number from here.

• gold — human-produced Chittagonian from a published, peer-reviewed corpus.
• silver — auto-derived: a label transferred from Bangla, or text produced by the transducer. Valid for relative model comparison; not a human-validated test set, and not a basis for absolute claims. Native-speaker validation is still outstanding.

Every item also carries a dialectness score in [0,1], so results can be reported by difficulty instead of collapsed into one aggregate. That matters: on the translation track a rule-based transducer improves as dialectness rises (40.1 → 66.5 chrF) while neural models degrade — an averaged number hides the entire effect.

The corpus

Dialect	Train	Val	Test	Total
Chittagonian	10,382	731	797	11,910
Sylheti	5,793	460	583	6,836
Barishal	4,112	373	480	4,965
Noakhali	3,581	336	479	4,396
Mymensingh	3,581	336	479	4,396
Rajshahi	1,707	87	104	1,898

Splits are assigned by hashing normalised text, and Vashantor's published splits are honoured where they exist.

Deduplicate before you combine these sources

23,617 raw records became 15,288 after deduplication — 35.3% were cross-source duplicates. Sentences from Vashantor's test split were present in ChatgaiyyaAlap's training data. Anyone merging these corpora without deduplicating is reporting inflated numbers.

Files

File	What
bench/*.jsonl	the 8 benchmark tasks
bench/bench_manifest.json	per-task counts, evidence grade, strata
corpus/parallel.jsonl	Bangla ↔ dialect pairs, 6 dialects, with splits
corpus/lexicon.json	word- and clause-level Bangla → Chittagonian mappings
artifacts/rules_ctg.json	656 mined transduction rules with held-out precision
artifacts/spelling_norm.json	orthographic variant → canonical form

Mined phonology

The rules were recovered automatically by character alignment over 10,382 sentence pairs — no linguistic rules were hand-written:

ক→গ (করা→গরা) · ক→হ (কথা→হতা) · প→ফ (পানি→ফানি) · খ→হ (খাবো→হাইয়্যুম) · হ→অই (হবে→অইব)

Validated against held-out text: standard Bangla scores 29.27 chrF against real Chittagonian, and the transducer 52.90. Note the honest figure — rules alone reach 35.15, and that is what applies to unseen vocabulary, because the lexicon short-circuits words it already knows.

Sources

Source	Licence	Via
ChatgaiyyaAlap	CC BY 4.0	Mendeley wtms9xbkkw
Vashantor	CC BY 4.0	Mendeley bj5jgk878b
ONUBAD	CC BY 4.0	Mendeley 6ft99kf89b
BD-Dialect	CC BY 4.0	Mendeley k769s4vk5z
Vulgar Lexicon	Apache-2.0	HF kit-nlp/Vulgar_Lexicon_of_Chittagonian_Dialect_of_Bangla_or_Bengali
titulm-bangla-corpus	CC BY 4.0	HF hishab/titulm-bangla-corpus
bengali_sentiment	see dataset card	HF mHossain/bengali_sentiment

Nothing here is scraped. Every source is an open-licensed academic release.

Citation

@software{chatgaiyyabench,
  title  = {ChatgaiyyaBench: A Multi-Task Benchmark for Chittagonian},
  author = {Asifuzzaman, A. K. M.},
  year   = {2026},
  url    = {https://github.com/A-K-M-Asifuzzaman/Chatgaiya-AI}
}