BanglaLLMBanglaLLM logo mark
Sign inJoin in

Help build a Bangla language model from Bangladesh's own Bangla

AI still understands little of the Bangla people speak across Bangladesh. Answer short questions in your own words and dialect. People from your region check the answers, and the accepted ones teach existing open weight models our Bangla.

  • About a minute per task
  • Write it, then say it aloud
  • Free and voluntary
  • Others see only your ID
Join in now See an example task
HOW IT WORKS
  1. 1 · Instruction

    Write “How are you?” the way you say it at home.

  2. 2 · You write
    তুমি কেমন আছ?Standard
  3. 3 · Checks
    Auto-check Reviewer 1 Reviewer 2
  4. 4 · Dataset row
    {
      "instruction": "Write “How are you?” the way you say it at home.",
      "response": "তুমি কেমন আছ?",
      "dialect": "bn-std",
      "status": "accepted"
    }
  5. 5 · AI learns
    GemmaQwen
    Existing open weight models, improved. Not built from scratch.
—volunteers
—contributions submitted
—verified examples in the dataset
—local dialects, one goal

What you will do

Short writing tasks, in the language of your home

Pick a task type to see an example. Write regional words the way you say them; there are no wrong spellings.

The instruction you see

Answer the question below in writing, the way you would explain it to someone at home.

শীতকালে আপনার এলাকায় কোন কোন পিঠা বানানো হয়?

An example answer

আমাদের এদিকে শীতে ভাপা আর চিতই পিঠা বেশি হয়। খেজুরের গুড় দিয়ে খাই।

A short answer from what you know yourself.

These are examples. You answer in the dialect you chose, in Bangla script. Each task takes about a minute and you can stop whenever you like.

How it works

Five layers, from your words to a Bangla model

1 / 5

Layer 1 · Input

You write and speak, the way you talk at home

Each task shows one instruction: a question, a situation, or a sentence to say your way. You type a short answer in Bangla script, then read sentences aloud and record them.

  • About one minute per task
  • For recording: a microphone and a quiet place
  • Your own words: no copying, no AI text
Answer

শীতকালে আপনার এলাকায় কোন কোন পিঠা বানানো হয়?

আমাদের এদিকে শীতে ভাপা আর চিতই পিঠা বেশি হয়। খেজুরের গুড় দিয়ে খাই।
Submit

Layer 2 · Checks

Automatic checks catch the obvious problems

Within seconds the system checks the basics, so reviewers only spend time on real answers.

  • Written in Bangla script
  • Long enough to be useful
  • Not a repeat of an earlier answer
  • Recordings are not silent or too noisy
  • Bangla script
  • Length
  • Not a duplicate
  • Really rewritten, for rewrite tasks

Passed: sent on for review

Layer 3 · Review

People from your own dialect confirm it

Reviewers who speak your dialect read your answer and decide whether it is accepted.

  • Reviewers see only an ID and an avatar, never your name
  • Experienced contributors can become reviewers too
  • Nothing is accepted without a person approving it

আমাদের এদিকে শীতে ভাপা আর চিতই পিঠা বেশি হয়। খেজুরের গুড় দিয়ে খাই।

  • Participant avatarReviewer 1same dialect Approve
  • Participant avatarReviewer 2same dialect Approve

2 of 2 approvals: accepted

Layer 4 · Dataset

Accepted answers become training examples

Each accepted answer is stored as an instruction and its response, labelled with the dialect. Names and emails are never part of it.

  • Saved in the format AI models learn from
  • A test set is held back, split by person
  • Withdraw consent and your rows are removed
{
  "messages": [
    { "role": "user",
      "content": "শীতকালে আপনার এলাকায় কোন কোন পিঠা বানানো হয়?" },
    { "role": "assistant",
      "content": "আমাদের এদিকে শীতে ভাপা আর চিতই পিঠা বেশি হয়। খেজুরের গুড় দিয়ে খাই।" }
  ],
  "dialect": "bn-std",
  "split": "train"
}
train 90%devtest

Layer 5 · AI learns

Existing open weight models learn our Bangla

The dataset improves existing open weight models, for example Gemma and Qwen. No new model is built from scratch.

  • Writing, so models read and write our dialects
  • Voice, so models understand and speak them, with your consent
  • The dataset may be shared openly for research
Existing open weight modelGemma · Qwenalready exists
+
Our datasetRegional Banglareviewed by speakers
The goalAn open weight model that reads and writes the way we speak
Swipe sideways to see every layer

What this platform is for

A dataset that improves existing AI models, not a new model

  • Written answers and voice recordings
  • Checked by people
  • Improves existing open weight models
  • For example Gemma and Qwen

What we are doing

  • You answer short tasks in writing and record your voice reading sentences aloud. Each task and your answer or recording become one example in the dataset.
  • The dataset helps existing open weight models, for example Gemma and Qwen, understand, write and speak Bangladeshi Bangla better.
  • Only answers that pass the automatic checks and are accepted by reviewers go into the dataset.

What we are not doing

  • We are not building a new AI model from scratch.
  • This is not a chatbot or an app to chat with.
  • We do not use your voice to imitate you.

Writing and voice

Writing

Written answers, conversations and sentences in your dialect. They teach AI to read and write our Bangla.

Voice

Recordings of you reading sentences aloud and speaking your answers, so AI can understand and speak our Bangla. You agree to recording in the consent form first.

Dialects across Bangladesh

The words of every region, the voice of every person

Choose the dialect you speak at home, or more than one. Each has its own goal. Questions also cover local food, festivals and sayings.

Your trust, our responsibility

Your words stay yours

You choose to take part

You agree before you start, knowing what is collected and why.

Others see only your ID

Only you and the project team see your name; everyone else sees a random ID and your avatar.

Leave any time

Withdraw your consent and your recordings are deleted, your answers leave all later datasets and your name is deleted.

Free and voluntary

It costs nothing. Give as much or as little time as you like.

“মোদের গরব, মোদের আশা,
আ মরি বাংলা ভাষা”
Our pride, our hope: O, my beloved Bangla language.— Atulprasad Sen

In 1952 we stood up for our mother tongue. Now let's carry it into tomorrow's technology. Give a few minutes, and give your language a future.

FAQ

What you may want to know

Does it cost anything?

No. Taking part is free and voluntary.

How much time do I need?

As much as you like. A task takes about a minute.

What do I need?

A phone or computer with internet and a Bangla keyboard. To record your voice, a microphone (your phone has one) and a quiet place. There is no app to install.

How should I spell regional words?

Write them the way you say them, in Bangla letters. There are no wrong spellings here.

What happens to my writing?

Your writing and recordings become part of a dataset that helps existing open weight models, such as Gemma and Qwen, understand and speak our Bangla, including speech-to-text and text-to-speech. It may be used for research and shared openly. The full terms are in the consent form when you join.

Will you record my voice?

Yes, in the voice tasks: you read sentences aloud or say your answer. A recording of your voice can identify you, so you agree to it in the consent form before you start. If you withdraw, your recordings are deleted, and your voice is never used to imitate you.

Can I use ChatGPT or copy from the internet?

No. The data is only useful if real people write it in their own words. Reviewers turn down copied or AI-written answers.

My dialect isn't listed. Can I still join?

Yes. Choose Standard Bangla or the closest dialect, and write the way you really speak.

Give ten minutes, give your language a future

Every answer you write helps AI learn the Bangla we really speak.

Start today
BanglaLLMBanglaLLM logo mark

My language, my AI

A data-collection platform. Your writing and voice recordings become a dataset that helps existing open weight models (for example Gemma and Qwen) understand and speak Bangladeshi Bangla. We are not building a new model from scratch.

The project

AboutHow it worksDialectsFAQ

Take part

Join inSign inSee an example task

Legal

Privacy PolicyTerms and Conditions
© 2026 BanglaLLM. A free, volunteer project.