Project Overview

This repository documents a practical, artifact-driven pipeline to build and deploy an on-device farmer advisory model using the KCC dataset. The pipeline comprises two main stages:

• Fine-tuning:

  • Dataset: "Farmers Call Query Data". A cleaning step was applied to remove all null and empty rows; the corresponding code is included in this repository.

    This was generated using data from data.gov.in, an open data platform by Govt. of India.

  • Base model: "unsloth/Qwen2.5-1.5B-Instruct" — chosen for its strong instruction-following performance and suitability for edge deployment.

• Conversion & Packaging:

  • The fine-tuned model was converted to LiteRT format (formerly TFLite) using the "AI Edge Torch" toolkit.
  • The resulting LiteRT model was packaged into LiteRT-LM format using utilities from the "LiteRT-LM" project. The final packaged model artifact is available on Hugging Face at https://doi.org/10.57967/hf/7392.

This pipeline closely mirrors the engineering principles and reproducibility focus of the prior Gemma-3n work, now adapted to the Qwen2.5 family to enable efficient on-device inference while preserving domain-specific relevance for advisory tasks among Indian smallholder farmers.