Share your details to access gemma3-12b-kcc-advisory

This model is a fine-tune trained on Kisan Call Centre (KCC) farmer query/answer data, published under the Government Open Data License – India (GODL). We gate access to understand who is using it. Your country and region (based on approximate Internet address) will be shared with the model owner.

Indic Crop-Advisory Gemma 3 12B (LoRA)

A LoRA adapter for unsloth/gemma-3-12b-it (base: google/gemma-3-12b-it), fine-tuned to answer Indian farmers' crop-advisory questions in the style of Kisan Call Centre (KCC) responses, in English and 10 Indian languages.

The adapter was trained on the instruction-tuned checkpoint, not the pretrained one. The Hub's model tree also lists google/gemma-3-12b-pt because that is the pretrained model both -it checkpoints were themselves tuned from; it is not a base for this adapter.

This is a text-only fine-tune. The base checkpoint is vision-capable, but the training data has no image rows, so the vision tower / projector were never exercised or updated — treat this adapter as language-model-only, not as a validated vision fine-tune.

Intended use

Draft/assistive crop-advisory answers (pest and disease queries, cultivation practices, fertilizer and irrigation guidance, market/scheme questions) for Indian farmers, in their own language and script. Intended as a decision-support aid reviewed by a human (extension officer, agronomist), not as a standalone source of truth for chemical dosages, pesticide selection, or any safety-critical recommendation.

Evaluation

Scored on Indic-KCC-Agri-Advisory-Benchmark (500 held-out KCC questions × 11 languages), LLM-judge rubric (Qwen3.6-35B-A3B judge), 1–5 scale, mean over all 11 languages:

model	params	correctness	naturalness	groundedness	safety	parse_ok
gemma3-12b-kcc-advisory (this model)	12B + LoRA r=32	3.44	4.13	3.93	4.54	0.98
gemma-3-27b-it (base, 2x the params)	27B	3.26	3.78	4.16	4.76	0.98
google/gemma-3-12b-it (base, unmodified)	12B	2.30	3.22	3.44	4.84	0.99

Highlights for a showcase framing:

+1.14 correctness / +0.91 naturalness / +0.49 groundedness over the unmodified google/gemma-3-12b-it base model it was fine-tuned from.
Beats gemma-3-27b-it — a base model at more than 2x the parameter count — on correctness and naturalness, despite being smaller and only LoRA-adapted, not full fine-tuned.

Data attribution: Government of India, Ministry of Agriculture & Farmers Welfare, Kisan Call Centre (KCC) — Transcripts of Farmers' Queries and Answers, data.gov.in. Published under the Government Open Data License – India (GODL). Raw KCC call-centre answers are unedited historical responses, not reviewed or graded by an agronomist — see Limitations.

How to use

With PEFT + Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "unsloth/gemma-3-12b-it"   # or google/gemma-3-12b-it
adapter_id = "sthanika-ai/gemma3-12b-kcc-advisory"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)

messages = [{"role": "user", "content": "मेरी टमाटर की फसल में पत्तियां पीली पड़ रही हैं, क्या करें?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Serving with vLLM (LoRA adapter, not merged)

vllm serve unsloth/gemma-3-12b-it \
  --enable-lora \
  --lora-modules kcc-advisor=<local-path-or-repo-id>

Limitations

Not a validated source for pesticide/fertilizer dosage or safety-critical advice — always have a qualified agronomist or extension officer verify recommendations before they're acted on.
Trained on machine-translated + reviewer-corrected text; translation artifacts may remain in lower-resource language/script combinations.
Vision inputs are untested — do not rely on this adapter for image-based (e.g. pest/disease photo) queries.
An early-stopped checkpoint by design (see Evaluation) — published because it benchmarks best, not because training "finished" here.

License

This is a fine-tune of Gemma 3 and is distributed under the Gemma Terms of Use and the Gemma Prohibited Use Policy. By using this adapter you agree to those terms. Built with Gemma. See the accompanying NOTICE file (required by the Gemma Terms of Use) and the Training data section above for the KCC dataset's GODL attribution.

This repository does not itself grant any rights to the Gemma base model beyond what Google's terms already grant, and nothing here implies endorsement by or affiliation with Google.

Framework versions

TRL: 0.24.0
Transformers: 5.5.0
PyTorch: 2.11.0+cu128
Datasets: 4.3.0
Tokenizers: 0.22.2
PEFT: 0.20.0

Citation

@misc{gemma3_12b_kcc_advisory,
  title        = {Indic Crop-Advisory Gemma 3 12B (LoRA)},
  author       = {sthanika-ai},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/sthanika-ai/gemma3-12b-kcc-advisory},
  note         = {LoRA fine-tune of google/gemma-3-12b-it on Kisan Call Centre (KCC) crop-advisory Q\&A}
}

This model was evaluated against, and its training data is a distinct derivative of the same source as, the following dataset:

@dataset{indic_kcc_agri_advisory_benchmark,
  title        = {Indic-KCC-Agri-Advisory-Benchmark},
  author       = {sthanika-ai},
  year         = {2026},
  version      = {1.0},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/sthanika-ai/Indic-KCC-Agri-Advisory-Benchmark},
  note         = {Derived from Kisan Call Centre (KCC) transcripts, Government of India, licensed GODL-India; see DATA_LICENSE.md in that dataset's repository}
}

Downloads last month: 18

Model tree for sthanika-ai/gemma3-12b-kcc-advisory

Base model

google/gemma-3-12b-pt

Finetuned

unsloth/gemma-3-12b-it

Adapter

(74)

this model

Evaluation results

correctness on Indic-KCC-Agri-Advisory-Benchmark
self-reported

3.440
naturalness on Indic-KCC-Agri-Advisory-Benchmark
self-reported

4.130
groundedness on Indic-KCC-Agri-Advisory-Benchmark
self-reported

3.930
safety on Indic-KCC-Agri-Advisory-Benchmark
self-reported

4.540