Back to DRISHTI-Bn case study

Research documentation

DRISHTI-Bn

Working documentation for an undergraduate thesis framework studying flood-damage classification and Bengali caption generation.

ViT-B/16BanglaBERTLoRACrisisMMDBengali AI

Status and scope

These are transparent working drafts. Final metrics, exact implementation details, dataset license information and publication status will be added only after verification from the code, experiment logs and official sources.

Architecture

ViT-B/16, BanglaBERT and LoRA pipeline draft.

ARCHITECTURE.md

Model card

Intended use, limitations and responsible-use guidance.

MODEL_CARD.md

Dataset notes

CrisisMMD context, Bengali annotation and preparation checklist.

DATASET.md

Reproducibility

Environment, training and evaluation checklist.

REPRODUCIBILITY.md

Ethics

Translation bias, privacy, safety and accountability notes.

ETHICS.md

Citation

Draft software citation metadata for the project.

CITATION.cff

Architecture overview

The documented pipeline combines a ViT-B/16 vision component and BanglaBERT language component, with LoRA-based parameter-efficient adaptation for two target tasks:

  • Flood damage classification.
  • Bengali caption generation.

The exact fusion, trainable layers and decoder behavior remain implementation-verification items rather than unsupported claims.

Dataset and annotation

The current project materials document 3,755 CrisisMMD flood images and machine-translated Bengali annotations.

  • Official source, citation and license: TBD.
  • Train, validation and test counts: TBD.
  • Cleaning and translation protocol: to be verified.

Responsible use

This is a research prototype, not an emergency-response authority or verified field-reporting service.

  • Human review is required for every operational interpretation.
  • Outputs must not drive life-critical decisions alone.
  • Machine-translated Bengali may contain language and cultural errors.

Reproducibility status

Reproducibility instructions are prepared as a checklist and will be completed from the final experiment configuration.

  • Python and dependency versions: TBD.
  • Preprocessing, training and evaluation commands: TBD.
  • Final evaluation metrics and qualitative examples: TBD.

Next verification steps

  1. 1. Verify the architecture against training and inference code.
  2. 2. Freeze and record the dataset split and evaluation protocol.
  3. 3. Re-run experiments and publish final metrics.
  4. 4. Confirm dataset license and citation before release.
  5. 5. Add supervisor details only with permission.
  6. 6. Publish the verified documents in the DRISHTI-Bn repository.
View the DRISHTI-Bn GitHub repository