maya-bn — real Bangla data for Laya typed decisions

655,241 real, native-Bangla records (304,552 labeled, 350,689 raw-for-labeling) in Laya's typed-decisions format.

Why this dataset exists

Laya is a non-autoregressive System-1 decision model: it takes a state (a text, email, ticket, review or JSON document) plus a set of typed questions — choice, noul (yes/no) and score (ordinal) — and returns a calibrated probability distribution over the answers in a single forward pass.

Laya's public checkpoints and its typed-decisions benchmark are English-centric. The multilingual checkpoint (laya-multilingual, mmBERT-base) is the right backbone for Bangla, but it is near chance on typed decisions before fine-tuning. This dataset exists to give Bangla that fine-tuning signal — and to serve the same role for comparable decision models such as TypeSafe Jev.

It is built real-data-first: instead of asking a teacher model to invent Bangla training cases, we collected existing, natively-Bangla, license-clean datasets, mapped their original human labels onto Laya's typed questions, and added a raw real-text pool for the domains where no labeled Bangla data exists.

Non-commercial superset available. Additional Bangla sources whose licences permit non-commercial use only are published separately at swadhinbiswas/maya-bn-nc (CC-BY-NC-4.0, 665,121 records). This repo stays permissive (CC-BY-4.0).

What it is for

• Fine-tuning convaiinnovations/laya / laya-multilingual (or any encoder) for Bangla typed decisions (classification → choice, yes/no → noul, ordinal rating → score).
• Benchmarking Bangla decision behaviour and calibration.
• A base for later synthetic gap-filling (Task 5 of the Maya spec) without re-inventing the parts of Bangla coverage that already exist.

Schema (Laya typed-decisions, drop-in)

column	meaning
id	stable record id
workflow	Bangla domain / task, e.g. sms_smishing, news_topic
split	train / validation / test
state	JSON string: the document the model must read
questions	JSON string: {name: {type, instructions, criteria}}
gold	JSON string: {name: {type, label, probabilities, confidence}}
factors	JSON string (empty for this data)
label_agreement	JSON string (empty; single-annotator gold)
n_questions	number of typed questions on the record

gold.probabilities are one-hot over the source's human label. Soft labels / temperature calibration are intentionally not applied here — that is Task 5.

The raw_state_pool config holds Type B: authentic documents with an empty questions/gold and needs_labeling = true. They are inputs for teacher labeling + human verification, not finished training labels.

Coverage

workflow	records	questions
news_document	257,471	—
nli	128,044	relationship
news_topic	60,829	category
spoken_sentence	53,283	—
health_document	32,478	—
hate_speech_binary	31,219	is_hate
social_sentiment	21,110	sentiment, sentiment_score
code_mixed_sentiment	19,721	sentiment
voice_assistant_intent	15,886	intent, scenario
youtube_review_sentiment	9,743	is_positive
financial_news_document	7,457	—
political_news_sentiment	5,826	sentiment, tone_score
sms_smishing	5,066	message_type
hate_speech	3,157	hate_type
review_sentiment	2,687	sentiment, sentiment_score
social_reasoning	1,264	address_term, answer

Full record-level provenance is in metadata/provenance_ledger.parquet (source, URL, license, collection date/method per record). The considered- sources inventory, including every rejected source and why, is in metadata/source_inventory.csv.

Sources (all natively-Bangla; license-clean by construction)

source	records
zabir-nabil/bangla_newspaper_dataset	257,471
nahid-hub/BLUGE-bengali-nli	128,044
nahid-hub/BLUGE-bengali-news-classification	60,829
bengaliAI/CommonVoiceBangla	53,283
faisal4590aziz/bangla-health-related-paraphrased-dataset	32,478
Yanjo/bengali-hate-speech-dataset	27,853
nahid-hub/BLUGE-bengali-sentiment-classification	21,110
asfaqur/bn_code_mix_sentiment_dataset	19,721
AmazonScience/massive	15,886
Akash190104/bengali_sentiment_analysis	9,743
ashtrayAI/Bangla_Financial_news_articles_Dataset	7,457
Mukaffi28/Motamot_Bengali_Political_Sentiment_Analysis	5,826
shariul-islam/bengali-sms-smishing-dataset	5,066
Lots-of-LoRAs/task1490_bengali_personal_hate_speech_binary_classification	3,364
rezacsedu/bn_hate_speech	3,157
Lots-of-LoRAs/task1496_bengali_reviews_sentiment_classification	1,358
Lots-of-LoRAs/task1497_bengali_book_reviews_sentiment_classification	1,329
sijantanvir/BanglaSocialBench	1,264
Lots-of-LoRAs/task1491_bengali_political_hate_speech_binary_classification	1
Lots-of-LoRAs/task1493_bengali_geopolitical_hate_speech_binary_classification	1

License mix of the pooled records

license	records
mit	291,241
cc-by-4.0	269,354
cc0-1.0	60,740
apache-2.0	33,906

The aggregate is released under CC-BY-4.0; every underlying component is CC0, CC-BY-2.5/4.0, MIT or Apache-2.0 (no non-commercial or share-alike components). Attribution is preserved per record in the provenance ledger.

Processing pipeline (reproducible)

1. Collect — every source downloaded from the Hub with huggingface_hub (scripts/03_fetch.py); no scraping, no paywalls, no shadow libraries.
2. Anonymise — emails, Bangladeshi phone numbers, long digit runs (NID / account / order numbers) and URL paths are masked before anything is stored as a state.
3. Filter — normalise (NFKC), strip HTML tags, trim repeat-char runs, drop non-Bangla-script rows from standard-register sources, drop boilerplate, apply per-workflow length bounds; re-tag mixed/romanised text as mixed.
4. Excerpt — newspaper-sourced raw text is reduced to the headline plus a short lead (copyright hygiene); no full article bodies are published.
5. Deduplicate — cross-source exact-hash and MinHash + LSH near-duplicate removal (n-gram Jaccard ≥ 0.75), so the same underlying post/article does not enter twice from two datasets.
6. Convert — original labels mapped to typed questions (see per-workflow definitions above); romanised Banglish kept as its own register, not silently mixed with native script.
7. Verify — automated QA (scripts/05_qa.py) re-scans for residual PII and emits a human checkpoint sample before release.

Honest limitations

• Not native for MASSIVE. voice_assistant_intent utterances were machine-translated from English then human-verified (authenticity = mt_from_english). They are included because Laya's own benchmark uses MASSIVE; treat them as translationese.
• High-cardinality choices. The MASSIVE intent question has 60 options, which is Laya's documented weak spot; scenario (18) is the easier companion question. Consider head_max_len tuning or routing.
• score (ordinal) coverage is derived, not native. No collected source provides native ordinal ratings. The score questions present map the 3-class sentiment valence (negative < neutral < positive) and binary review/news tone onto an ordinal scale; treat native ordinal Bangla as an open gap.
• Hate/toxicity data contains offensive language by design. It is included for moderation-style decision tasks, not as neutral text; do not repurpose it as a general LM corpus.
• Hard labels only. No annotator distributions; confidence is one-hot. Do not read model confidence from this file without calibration.
• Domain gaps with no real Bangla source (synthetic-only in the main spec): invoice/accounts-payable processing, security-incident triage, IT helpdesk triage, and agent-trace observability.
• Not a factual-knowledge corpus. News/review/health text is present for register diversity, not for factual use.

Intended use

Fine-tuning and evaluating Bangla decision/classification models (Laya, laya-multilingual, TypeSafe Jev or similar). Not for generating text, and not as a source of ground truth about people or events.

Citation / provenance

Built from the public sources listed above; cite each original dataset (see metadata/provenance_ledger.parquet for the exact source per record). Pipeline and collection ledger generated 2026-09-22.