kcc-krishi-qwen3-1.7b/README.md

---
language:
- en
license: other
base_model:
- Qwen/Qwen3-1.7B
library_name: transformers
tags:
- agriculture
- india
- kisan-call-centre
- farmer-advisory
- qwen3
- qlora
- rag
- safety-routing
- jetson
- offline-inference
pipeline_tag: text-generation
---

# KCC-Krishi Qwen3-1.7B

KCC-Krishi Qwen3-1.7B is a compact English-primary agricultural advisory model fine-tuned from `Qwen/Qwen3-1.7B` using QLoRA on a routing-aware Kisan Call Centre research corpus.

The model is designed for a regional offline farmer-assistance architecture:

```text
regional speech
→ local ASR
→ translation to English
→ runtime guard
→ regional KCC RAG retrieval
→ KCC-Krishi Qwen3-1.7B
→ agrochemical exactness validation
→ translation to farmer language
→ local TTS
```

## Important status

This model is a research and controlled-prototyping artifact.

It is not:

- field-validated agronomic ground truth;
- an unrestricted agricultural authority;
- a replacement for agronomists;
- a replacement for current official advisories;
- a replacement for product-label verification;
- a replacement for farmer feedback or field validation;
- a general-purpose assistant for arbitrary non-agricultural questions.

## Training details

Base model:

```text
Qwen/Qwen3-1.7B
```

Training method:

```text
QLoRA / LoRA
```

Training stack:

```text
Hugging Face Transformers
TRL SFTTrainer
PEFT
bitsandbytes
```

Training hardware:

```text
Azure Standard_NC8as_T4_v3
NVIDIA Tesla T4 16GB
8 vCPUs
56 GiB system RAM
```

Final training configuration:

- training examples: 365,037;
- validation examples used during training: 5,000;
- one complete training epoch;
- optimizer: `paged_adamw_32bit`;
- learning rate: `1e-5`;
- warmup ratio: `0.05`;
- maximum gradient norm: `0.3`;
- QLoRA compute dtype: `float16`;
- Trainer FP16: disabled;
- Trainer BF16: disabled;
- packing: disabled;
- assistant-only loss: disabled;
- maximum sequence length: 1,024;
- manual Qwen ChatML rendering;
- hard-stop monitoring enabled.

Training duration:

```text
approximately 85.8 hours
```

The full training run completed without NaN loss, Inf loss, loss explosion, CUDA out-of-memory failure, or hard-stop activation.

## Final model selection

The selected adapter originated from the exact complete-epoch checkpoint:

```text
checkpoint-22815
```

The selected adapter was compared against the validation-selected checkpoint-18000 using normal RAG, OOD, and adversarial behavioral probes.

The selected candidate performed better on several unsafe or adversarial cases, including:

- pesticide mixing;
- excessive fertilizer use;
- double pesticide dose requests;
- harmful pesticide framing;
- out-of-distribution agricultural questions.

The model was then merged with the Qwen base model using float16 safe serialization.

## Merged model files

The merged release includes:

- `model.safetensors`;
- `config.json`;
- `generation_config.json`;
- `tokenizer.json`;
- `tokenizer_config.json`;
- `chat_template.jinja`;
- `merge_summary.json`;
- `SHA256SUMS.txt`.

## Required runtime architecture

The model should not be deployed alone.

The intended runtime is:

```text
out-of-scope guard
→ current-information guard
→ regional KCC RAG retrieval
→ fine-tuned Qwen model
→ agrochemical exactness guard
→ final safety validation
```

The runtime guard is required because the model is not intended to answer arbitrary non-agricultural questions. The guard should block or safely handle general mathematics, politics, fictional premises, nonsense prompts, and unrelated requests.

## Agrochemical exactness requirement

The model may alter exact product names, formulations, percentages, or dose expressions when generating from retrieved evidence.

Therefore, every generated answer containing pesticide, herbicide, fungicide, fertilizer, formulation, percentage, or dose-like information must be checked against verified retrieved evidence.

If the generated agrochemical information is not supported exactly by retrieved evidence, the runtime should replace it with generic wording such as:

```text
Use only a locally recommended product at the current label dose. Confirm the crop, pest, crop stage, formulation, and local agricultural recommendation before application.
```

## Intended use

Suitable uses include:

- agricultural RAG research;
- KCC evidence-grounding experiments;
- supervised fine-tuning research;
- offline farmer-assistant prototypes;
- Jetson and mobile inference experiments;
- safety-routing research;
- multilingual voice-assistant integration through external ASR, translation, and TTS services.

## Unsuitable use

Do not use this model alone for:

- direct pesticide or fertilizer recommendations;
- autonomous disease diagnosis;
- current weather forecasting;
- current mandi-price reporting;
- government-scheme status;
- financial or medical decisions;
- unrestricted general-purpose chat;
- unattended field deployment.

## Dataset

The final model was trained using the KCC-Krishi routing-aware corpus.

Dataset repository:

```text
uralstech/kcc-krishi-rag-sft-advisory-corpus
```

The Punjab/Wheat GPT-5.5 normalization demonstration subset was not used for final Qwen training.

## Dataset attribution

The underlying training evidence is derived from:

**Kisan Call Centre (KCC) Transcripts of Farmers Queries and Answers**

Open Government Data Platform India / Data.gov.in
Ministry of Agriculture and Farmers Welfare, Government of India
Accessed through AIKosh/Data.gov.in

The KCC-Krishi processing, routing, training, packaging, and runtime safeguards were created by:

```text
KCC-Krishi Project Team
URAV ADVANCED LEARNING SYSTEMS PRIVATE LIMITED
```

## Base model and licensing

This model is derived from `Qwen/Qwen3-1.7B`.

The Qwen base model repository provides its own license and attribution terms. Users must review and comply with those terms before using or redistributing this derived model.

This model repository uses `license: other` because the release combines:

- the Qwen base model;
- KCC-derived training evidence;
- project-specific fine-tuning;
- project-specific model packaging;
- runtime safety requirements;
- source and derived-data attribution obligations.

See the included `LICENSE.md` for the project release notes.

## Limitations

Known limitations include:

- possible source-data noise and query/evidence mismatch;
- historical or stale KCC records;
- lack of agronomist field validation;
- lack of broad farmer-feedback validation;
- lack of RLHF or preference tuning;
- possible model memorization of source-derived records;
- occasional terse or evidence-report-like responses;
- possible multilingual or translation artifacts;
- possible exact agrochemical-name or formulation drift without the runtime guard;
- poor behavior on arbitrary non-agricultural questions without the runtime guard.

## Related resources

Dataset:

```text
https://huggingface.co/datasets/uralstech/kcc-krishi-rag-sft-advisory-corpus
```

Existing Punjab/Wheat demonstration dataset:

```text
https://huggingface.co/datasets/uralstech/kcc-krishi-punjab-wheat-demo
```

Related preprint DOI:

```text
10.5281/zenodo.21040992
```

## Disclaimer

This model is released for research, reproducibility, RAG, SFT, and controlled offline-assistant prototyping.

Do not rely on this model as current or authoritative agricultural advice. Verify all consequential decisions with qualified agricultural experts, current official sources, current product labels, and applicable local regulations.