KisanSLM-GGUF

KisanSLM

KisanSLM is unsloth/Qwen3.5-2B fine-tuned (QLoRA, 4-bit) to answer Indian farmers' agricultural questions in the register of a real Kisan Call Centre (KCC) / Farm Tele Advisor (FTA) advisor — short, specific, actionable advice (product / dose / timing), instead of generic AI-assistant hedging.

Training code, the full dataset pipeline, and evaluation scripts are here: github.com/PrabhudattaPatra/KisanSLM

Files

Training data

~13,241 curated farmer-query / FTA-answer pairs, derived from India's Kisan Call Centre transcripts (data.gov.in, Government Open Data License – India). Built via a pipeline that pulled 400K+ raw records and cleaned, deduplicated, and quality-filtered them (rule-based + LLM-judge scoring) down to this set. Answers are kept verbatim from the real FTA response — nothing LLM-rewritten. Full methodology and stats: see the GitHub repo's DATASET_CARD.md.

Training method

• Base model: unsloth/Qwen3.5-2B
• Method: QLoRA (4-bit), via Unsloth Studio
• Hardware: Google Colab, T4 GPU (16GB VRAM)
• 11,916 training examples, 662 validation
• Exported to GGUF, Q4_K_M quantization

How to run (Ollama)

ollama pull hf.co/prabhu09/KisanSLM-GGUF:Q4_K_M
ollama cp hf.co/prabhu09/KisanSLM-GGUF:Q4_K_M kisanslm

The system prompt matters — set it via a Modelfile (ollama create kisanslm -f Modelfile):

FROM hf.co/prabhu09/KisanSLM-GGUF:Q4_K_M
SYSTEM """You are a Farm Tele Advisor (FTA) at India's Kisan Call Centre (KCC). Farmers call in with questions about crops, pests, diseases, soil, irrigation, weather, and government agricultural schemes. Give specific, locally-relevant, actionable advice in a direct, practical register -- the way a real KCC advisor would, not a generic AI assistant."""

ollama run kisanslm --think=false "My paddy leaves have brown spots, what should I do?"

Qwen3.5 has a built-in "thinking" reasoning phase that can loop indefinitely on this quantized 2B model. Always pass --think=false (CLI) or "think": false (API) — see the GitHub repo's FINETUNE.md for details.

Evaluation results and known limitations

Evaluated on 30 held-out test examples (never seen in training) against the real FTA answers. Full report/methodology in the GitHub repo.

• Style/register transfer: works well. Answers are short, direct, product+dose+timing style — close to indistinguishable from real FTA answers on plant-protection queries, the dataset's largest category.
• Factual grounding: a real limitation. On specific facts the model wasn't taught precisely enough to recall — market prices, phone numbers/contacts, exact pesticide product names — it confidently fabricates plausible-looking wrong answers rather than hedging. This is expected for a 2B model with no retrieval grounding; fixing it needs RAG against a live source, not more fine-tuning.

Practical guidance: trust this model's plant-protection / cultural practice / nutrient-management advice as a starting point (verify doses locally); do not trust prices, contact details, or exact product names without independent verification.

License

• Base model (unsloth/Qwen3.5-2B): Apache 2.0.
• Training data: Government Open Data License – India (GODL).