phoenix28/fasal-mitra-sft-v1

Fasal Mitra SFT Dataset v1

Fasal Mitra ("Farmer's Friend") is a supervised fine-tuning (SFT) dataset for training multilingual crop disease advisory AI models, specifically Gemma 4 E4B via Unsloth.

Dataset Summary

This dataset enables a vision-language model to diagnose crop diseases from photos and provide structured agricultural advisory in Assamese, Hindi, Bengali, and English — the primary languages of farmers in Northeast India (Assam, Meghalaya, Nagaland).

Split	Examples
Train	12,728
Validation	1,500
Test	1,500
Total	15,728

Each example = one crop image + one farmer query (in one of 4 languages) + one structured advisory response (generated by google/gemma-4-31B-it teacher model).

Data Sources

Source	Images	Proportion
PlantVillage (HuggingFace: mohanty/PlantVillage)	~2,380	59.5%
Multi-Crop Disease Dataset (Roboflow, YOLO format)	~960	24.0%
Paddy Disease Classification (Kaggle)	~456	11.4%
Pest_V2 (rice pest images, 10 classes)	~138	3.5%
Tea Leaf Dataset (Assam tea diseases)	~66	1.6%
Total unique images	~4,000	100%

4,000 images × 4 languages = ~16,000 training examples

Languages

Language	ISO Code	Script
English	en	Latin
Hindi	hi	Devanagari
Assamese	as	Bengali (Assamese)
Bengali	bn	Bengali

Translations generated by ai4bharat/indictrans2-en-indic-dist-200M.

Crops and Diseases Covered

Rice/Paddy: blast, bacterial_leaf_blight, brown_spot, bacterial_leaf_streak, bacterial_panicle_blight, tungro, hispa, dead_heart, downy_mildew, healthy

Rice Pests: asiatic_rice_borer, brown_plant_hopper, yellow_rice_borer, rice_leaf_roller, small_brown_plant_hopper, rice_leafhopper, rice_water_weevil, paddy_stem_maggot, rice_gall_midge, rice_leaf_caterpillar

Banana: sigatoka, moko_disease, panama_wilt, bract_mosaic_virus, cordana_leaf_spot, healthy

Cauliflower: black_rot, bacterial_spot_rot, downy_mildew, healthy

Chilli: anthracnose, leaf_curl, leaf_spot, whitefly, yellowing, healthy

Groundnut: early_leaf_spot, late_leaf_spot, rust, nutrition_deficiency, healthy

Radish: black_leaf_spot, downy_mildew, flea_beetle, mosaic_virus, healthy

Tea (Assam): blister_blight, red_spider_mite, grey_blight, red_rust (if available)

PlantVillage crops: Apple, Corn, Grape, Tomato, Potato, Pepper, Strawberry, Peach, Cherry, Squash, Blueberry, Orange, Soybean (26 disease classes)

Dataset Fields

{
    "id":             str,    # unique example ID e.g. "paddy_00042_hi"
    "image":          Image,  # PIL Image (RGB)
    "source_dataset": str,    # origin dataset name
    "crop":           str,    # crop type e.g. "rice", "banana"
    "disease_class":  str,    # disease label e.g. "blast", "healthy"
    "language":       str,    # "english" | "hindi" | "assamese" | "bengali"
    "query":          str,    # farmer question in target language
    "advisory":       str,    # structured advisory response in target language
    "messages_json":  str,    # Gemma 4 chat messages format (JSON string)
}

Usage

from datasets import load_dataset

ds = load_dataset("phoenix28/fasal-mitra-sft-v1")
example = ds["train"][0]

# For Unsloth/TRL fine-tuning, parse messages_json:
import json
messages = json.loads(example["messages_json"])
# messages[0] = {"role": "user", "content": [{"type":"image"}, {"type":"text","text":"..."}]}
# messages[1] = {"role": "assistant", "content": "**Disease:** ..."}

Advisory Format

Each advisory follows this 6-section structure (required by GRPO reward functions):

**Disease:** [scientific name] ([Hindi/Assamese local name])
**Severity Assessment:** [mild/moderate/severe — visual description]
**Immediate Treatment:**
- Apply [ICAR-approved chemical] at [dose per liter]
- Spray interval: Every [X] days, [Y] applications
**Spray Timing:** [morning/afternoon, weather conditions]
**Prevention Next Season:**
- Use resistant variety: [name]
- Seed treatment: [treatment]
**Consult:** [KVK guidance, ICAR toll-free 1800-180-1551]

Safety Constraints

• All treatment recommendations are ICAR-approved (Indian Council of Agricultural Research)
• Advisory responses were checked against the CPCB/CIB&RC banned pesticide list
• No banned pesticides (monocrotophos, endosulfan, methyl parathion, DDT, etc.) are recommended

Model Card

Fine-tuned model: phoenix28/fasal-gemma-4-e4b

Citation

@dataset{fasalmitra_sft_v1_2026,
  title     = {Fasal Mitra SFT Dataset v1 — Multilingual Crop Disease Advisory},
  author    = {phoenix28},
  year      = {2026},
  publisher = {HuggingFace Hub},
  url       = {https://huggingface.co/datasets/phoenix28/fasal-mitra-sft-v1},
  note      = {Created for Gemma 4 Good Hackathon 2026, Unsloth track}
}

Acknowledgements

• PlantVillage Dataset — S. Mohanty et al. (CC BY-SA 4.0)
• Paddy Doctor Dataset — ACM CODS-COMAD 2023
• AI4Bharat IndicTrans2 — translation model
• Google DeepMind Gemma 4 — teacher model for advisory generation
• iNaturalist — field observation photos