The Dataset

The current data preview of Bajjika is being released as part of the Project BoLI. This preview is a reflection of the full dataset and consists of the following -

Speech Recordings of 10 narrations in the language.
Transcriptions in IPA and native script(s).
Translations in English (which also act as prompts for the translation sentences).
Detailed speaker metadata, including their demographic, educational and linguistic profile.
Prompt in English and Hindi.

The full dataset contains the following -

Translations of a minimum of 1000 carefully selected sentences. These sentences are selected to represent diverse morphosyntactic categories generally found in Indian languages such as demonstratives, classifiers, TAM morphology, and different sentence structures such as transitives and ditransitives. These sets of sentences are based on the standardised questionnaires built by Linguists for writing the sketch grammar of any human language, thereby, representing almost full range of morphosyntactic properties exhibited in the language. Unlike other benchmarks which focus largely on lexical level evaluation (aka domains), this is the first benchmark dataset that evaluates model's performance on a range of morphosyntactic structures, even the most uncommon ones.
More than 600 narrative speeches collected across 8 domains. These are recorded by at least 2 speakers. Narrations, along with their prompts, can be used to evaluate AI models on a range of prompt-based tasks.
Along with transcriptions in IPA and other scripts and translation, all data is interlinearly glossed at morphemic level. This gives a word-by-word meaning and morphosyntactic information, thereby, enabling evaluation of models on their deep grammatical knowledge, ability for cross-linguistic comparison and generalisation and reasoning capacity and skills in language-related puzzles. This allows for evaluating the model's capacity on reasoning tasks beyond mathematical reasoning tasks as well as their capacity to arrive at typological generalisations.
In addition to the data and prompt itself, as mentioned earlier, we are also making available the detailed speaker metadata and prompt-level metadata viz domain, elicitation method, multilingual prompt, target grammatical category (for translation sentences), etc.
In accordance with our data governance policy, all contributors to the dataset, including those recording the dataset and those transcribing it, are named and listed as data contributors.

Dataset Preparation

The speech included as part of this dataset was recorded by native speakers using a data collection application on a mobile device (Karya or in-house app, Atekho), by translating sentences from English or Hindi to the target language.

The recorded speech was validated and then transcribed manually by trained linguists, working with the native speakers, following the guidelines for the project using MATra Lab, a part of the LiFE Suite Ecosystem, developed by Unreal Tece LLP.

Project BoLI Guidelines

About Bajjika

Bajjika belongs to the Indo-Aryan language family, specifically situated within the Eastern Zone (Bihari) subgroup. Taxonomically, its status is subject to ongoing classification debates: while it is assigned its own ISO 639-3 code (vjk) in some registries, trusted metadata also associates its Glottocode (bajj1234) with Maithili (mai), reflecting its transitional position and close structural relationship between Maithili and Bajjika.

Speakers

Due to the official practice of aggregating Bajjika speakers under Hindi in the Census of India, precise official demographic figures remain highly uncertain. However, academic assessments estimate the speaker population to be approximately 15 to 20 million (Pandey, 2023).

Distribution in India

Bajjika is primarily spoken in the north-central parts of the state of Bihar, India, chiefly within the Tirhut division. Its core linguistic region includes the districts of Muzaffarpur, Vaishali, Sitamarhi, Sheohar, and Samastipur, and it extends northward toward the Terai region of Nepal.

Major grammatical features

Phonology: Features a typical Indo-Aryan phonemic inventory, including a contrast between dental and retroflex consonants, and a distinction between aspirated and unaspirated stops. Vowel nasalization is phonemic, and vowel shortening or deletion rules apply in non-initial syllables.
Morphology: Exhibits a complex verb morphology characteristic of the Bihari group, featuring polypersonal agreement where verb inflections can index both the subject and the object, heavily modulated by levels of honorificity. Nouns generally employ postpositions for case marking rather than fusional case endings.
Syntax: Adheres to a default Subject-Object-Verb (SOV) word order. It is head-final, utilizing postpositions rather than prepositions, with modifiers preceding the noun heads they qualify.

Further reading

About Project BoLI

Project BoLI is the flagship project of UnReaL-TecE LLP, which aims to build high-quality datasets for benchmarking and evaluating different kinds of AI tasks, including speech-to-text, machine translation, grammatical analysis and reasoning tasks and prompt-based evaluation of LLMs. While the project aims to build these datasets for every Indian language and variety, the primary focus is on over 1300 underserved languages and both first and second language varieties of major, scheduled languages. The project's uniqueness is not just limited to the kind of benchmarking tasks it supports (including proposing some novel tasks) and also the kind of languages and communities it supports but also in its contextualisation and implementation of a unique data governance model (not yet implemented anywhere across the globe), which mandates that all datasets released by the project and their derivatives (including the models) are co-owned by the community members and all contributors of the project, and any permission to use the dataset is not a transfer of ownership but a revocable license to use it. The conditions under which the license could be revoked are clearly mentioned as part of the BoLI License. The complete details of the project, languages and communities supported till now, the quantum of data available till now, its data governance model and other relevant documents are all publicly accessible at the project website.

Ethical Considerations, Consent, IPR and Attribution

Project BoLI and this repository represents our commitment to not only fair remuneration to the speakers of the language but also to co-ownership and equal IPR to all the contributors who have built the dataset. We believe this is the first step to move away from the extractive data collection and use practices and ensure fairness in our treatment of the community members. As such, we have listed all speakers and transcribers as Contributors to the dataset (we insist that they are co-owners of the dataset, even though the HuggingFace platform does not provide us an explicit way of stating that) and they are further recognised as Speakers and Annotators of the dataset. This dataset is only licensed to other researchers for use in their research projects. More details about licensing and commercial use conditions are given in the License and Commercial Use sections.

Project BoLI - Data Governance Policy

BoLI Ethics Principle & Pledge

Project BoLI - Digital Consent Form

Project BoLI - Field Recording of Oral Consent

Project BoLI - TnC

Dataset Access

Full dataset for the language can be browsed and queried on our app. The results of the evaluation of different models will also be made available on the same link. If you would like to access the full dataset for your own use or would like to work with us in collecting more data for any language or variety, or collaborate with us in this initiative in any other way, please get in touch with us.

Direct relationships

Is record of 1

Is in language 2

Has transcription 1

Continue exploring