# IlaAI-v3 ๐ฑ
### Earth ยท Crop ยท Intelligence
**An open-source Agricultural AI for Bharat's Farmers**
[](https://huggingface.co/Ila-AI)
[](https://github.com/IlaAI)
[](LICENSE)
[]()
---
## ๐พ What is IlaAI?
**IlaAI** (เคเคฒเคพ โ Sanskrit for *"the earth that gives"*) is an open-source LLM series built specifically for Indian agriculture. Our mission is simple โ give every farmer in India access to expert agricultural advice, for free, in their own language.
> "For the hands that feed a billion ๐พ"
---
## ๐ What's New in v3?
IlaAI-v3 is our biggest leap yet โ trained on **real farmer data** from India's Kisan Call Center (KCC).
- ๐ **Best val loss ever โ 0.598**
- ๐พ **175,000+ real farmer Q&A pairs** from KCC
- ๐ฃ๏ธ **Telugu support** (experimental)
- โก **Fast responses** with `enable_thinking=False`
- ๐ฎ๐ณ **India-specific** โ real crops, real problems, real solutions
## ๐ How v3 Compares
| Feature | v1 | v1.1 | v3 |
|---------|-----|------|-----|
| Dataset | Synthetic 96K | Thinking 100K | **Real KCC 175K** |
| Val Loss | 0.695 | 0.821 | **0.598** |
| English | โ
Good | โ
Better | โ
Excellent |
| Hindi | โ Weak | โ ๏ธ Basic | โ
Good |
| Telugu | โ None | โ None | โ ๏ธ Experimental |
| Real Farmer Data | โ | โ | โ
|
| Thinking Disabled | โ | โ | โ
|
| Response Speed | Slow | Slow | **Fast** |
---
## ๐พ About the KCC Dataset
The **Kisan Call Center (KCC)** is a Government of India initiative launched in 2004. Farmers across India call a toll-free number (1800-180-1551) to get expert advice on agriculture. Every call is logged โ the farmer's question and the expert's answer.
This dataset contains **175,000+ real farmer questions** with expert answers, covering:
- Crop diseases and pest management
- Fertilizer and pesticide recommendations
- Irrigation and water management
- Government schemes and subsidies
- Market prices and crop selection
- Region-specific advice across all Indian states
Training on this data means IlaAI speaks like a real agricultural expert โ because it learned from real agricultural experts. ๐ฑ
---
## ๐ Quick Start
```python
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler
model, tokenizer = load("Ila-AI/IlaAI-v3")
messages = [
{"role": "system", "content": "You are IlaAI, an expert agricultural assistant for Indian farmers. Always respond in the same language the user writes in. Be concise, practical and helpful."},
{"role": "user", "content": "My wheat crop has yellow spots on leaves. What should I do?"}
]
text = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=False,
enable_thinking=False # Always add this for fast responses!
)
sampler = make_sampler(temp=0.3, top_p=0.9)
response = generate(model, tokenizer, prompt=text, max_tokens=500, sampler=sampler, verbose=True)
```
---
## โก Important โ Always Disable Thinking Mode
```python
text = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=False,
enable_thinking=False # Critical! Always add this.
)
```
Without `enable_thinking=False`, the model wastes tokens on internal reasoning and responses are slower and lower quality.
---
## ๐ Training Details
| Detail | Value |
|--------|-------|
| Base Model | IlaAI-v1.1 (Qwen3-4B) |
| Framework | MLX LoRA |
| Hardware | Apple M4 Mac Mini (24GB) |
| Dataset 1 | KCC 175K real farmer Q&A |
| Dataset 2 | KissanAI Thinking-climate-100k |
| Dataset 3 | 1,515 Telugu Q&A pairs |
| Training iters | 3,000 |
| LoRA rank | 8 |
| Final Val Loss | **0.598** |
| Peak Memory | 8.3 GB |
---
## ๐ฃ๏ธ Language Support
| Language | Status |
|----------|--------|
| English | โ
Excellent |
| Hindi | โ
Good |
| Telugu | โ ๏ธ Experimental |
| Tamil | ๐ Coming in v4 |
| Kannada | ๐ Coming in v4 |
| Marathi | ๐ Coming in v4 |
| Bengali | ๐ Coming in v4 |
| Gujarati | ๐ Coming in v4 |
| Punjabi | ๐ Coming in v4 |
---
## ๐ Model Evolution
| Version | Dataset | Val Loss | Status |
|---------|---------|----------|--------|
| v1 | KissanAI 96K synthetic | 0.695 | โ
Released |
| v1.1 | KissanAI 100K thinking | 0.821 | โ
Released |
| v3 | KCC 175K real + Telugu | **0.598** | โ
Released |
| v4 | Full 22+ Indian languages | TBD | ๐ Coming |
---
## ๐บ๏ธ Complete Roadmap
### โ
Phase 1 โ Foundation (Done)
- [x] IlaAI-v1 โ English agriculture advisory (96K synthetic rows)
- [x] IlaAI-v1.1 โ Improved English (100K thinking dataset)
- [x] IlaAI-v3 โ KCC real farmer data + Telugu experimental
### ๐ Phase 2 โ True Multilingual (v4)
- [ ] 22+ Indian languages with proper quality
- [ ] Real multilingual agriculture Q&A
- [ ] BPCC + AI4Bharat datasets
- [ ] Release IlaAI-v4
### ๐๏ธ Phase 3 โ Vision (IlaAI-Vision)
- [ ] Crop disease detection from photos
- [ ] Support all major Indian crops
- [ ] Combine text + vision into one model
### ๐ฑ Phase 4 โ Mobile App (Free Forever)
- [ ] Free Android app
- [ ] Free iOS app
- [ ] On-device inference (no internet needed)
- [ ] Voice input in Indian languages
- [ ] Available on Play Store & App Store
### ๐ Phase 5 โ Global
- [ ] Expand beyond India
- [ ] Support farmers worldwide โ Africa, Southeast Asia, South America
- [ ] Open API for developers
- [ ] WhatsApp bot integration
---
## ๐ค Contributing
We welcome contributions from developers, farmers, agronomists, and language experts!
- ๐ Add regional crop data
- ๐ Help with Indian language translations
- ๐ Report issues
- ๐ก Suggest features
---
## ๐ License
Apache 2.0 โ free to use, fine-tune, and build upon.
---
## ๐ Acknowledgements
- [KissanAI](https://huggingface.co/KissanAI) for open-sourcing Dhenu models and datasets
- [Kisan Call Center](https://data.gov.in) โ Government of India
- [Qwen Team](https://huggingface.co/Qwen) for Qwen3 base models
- [Apple MLX Team](https://github.com/ml-explore/mlx) for MLX framework
- Every Indian farmer who inspired this project ๐พ
---