A domain-specific language model fine-tuned for agricultural analysis tasks. This model provides structured insights on farm operations, investment opportunities, risk assessment, and regional agricultural profiling.

Project Scope

This release focuses on 8 major US agricultural counties as a proof-of-concept:

Fresno, CA (Tree nuts, Grapes)
Kern, CA (Almonds, Pistachios)
Lancaster, PA (Dairy, Corn)
Sioux, IA (Corn, Soybeans)
Yakima, WA (Apples, Hops)
Imperial, CA (Alfalfa, Lettuce)
Monterey, CA (Strawberries, Lettuce)
Deaf Smith, TX (Cattle, Wheat)

Upcoming releases will include:

Full US county coverage (3,000+ counties)
Specialized models for crop yield prediction
Harvest timing optimization
Soil health analysis
Weather impact assessment

Performance Benchmarks

Evaluated against base Phi-3 Mini on agricultural analysis tasks:

Content Coverage

Base Model    |████████████████████                    | 50.0%
Fine-Tuned    |█████████████████████████               | 63.3%  (+26.7%)

Structure Quality

Base Model    |██████████████████████                  | 56.7%
Fine-Tuned    |█████████████████████████████████       | 83.3%  (+47.1%)

Inference Speed

Base Model    |████████████████████████████████████████| 101.9s
Fine-Tuned    |████████████████████████                | 62.4s  (38.8% faster)

Summary Table

Metric	Base Model	Fine-Tuned	Improvement
Content Coverage	50.0%	63.3%	+26.7%
Structure Quality	56.7%	83.3%	+47.1%
Avg Inference Time	101.9s	62.4s	38.8% faster

Training Loss Curve

Epoch 1: ████████████████████████████████████████ Loss: 1.605 → 0.115
Epoch 2: ████████████                             Loss: 0.115 → 0.020
Epoch 3: ████                                     Loss: 0.020 → 0.019

The fine-tuned model shows significant improvement in generating domain-relevant content and producing well-structured analytical outputs.

Training Configuration

Parameter	Value
Base Model	microsoft/Phi-3-mini-4k-instruct
Method	LoRA (Low-Rank Adaptation)
LoRA Rank	64
LoRA Alpha	128
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Examples	900
Validation Examples	100
Epochs	3
Batch Size	16
Learning Rate	2e-4
Precision	BF16
Training Time	~51 minutes
Final Eval Loss	0.0186
Token Accuracy	99.3%

Hardware: NVIDIA DGX Spark (Blackwell GPU, 128GB RAM)

Use Cases

Farm investment analysis
Regional agricultural profiling
Risk factor identification
Technology adoption recommendations
Irrigation and water sustainability assessment
Crop expansion planning

Usage

python

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

# Load model
model_name = "microsoft/Phi-3-mini-4k-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

base_model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
model = PeftModel.from_pretrained(base_model, "sarathi-balakrishnan/phi3-agricultural-analyst")

# Create prompt
prompt = """You are an expert agricultural analyst. Analyze the following query using the provided data context.

### Query
What are the key investment opportunities in Fresno County agriculture?

### Data Context
County: Fresno, CA
Operators: 5,847
Average Farm Size: 423 acres
Irrigation Coverage: 89%
Revenue per Acre: $2,340

### Analysis"""

# Generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    inputs["input_ids"],
    max_new_tokens=500,
    temperature=0.7,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Output Format

The model generates structured analysis with:

Summary of the agricultural region
Key insights based on data patterns
Investment/growth opportunities
Risk factors and constraints
Confidence level

Limitations

Currently covers 8 US counties (expansion planned)
Training data is synthetically generated
Should not replace professional agricultural consulting
Performance may vary for regions outside training distribution

Technical Notes

Uses eager attention for Phi-3 compatibility
Optimized for BF16 inference on modern GPUs
LoRA adapters can be merged for deployment: model.merge_and_unload()

License

MIT License - Free for commercial and research use.

Contact

For questions about this model or collaboration on agricultural AI:

HuggingFace: @sarathi-balakrishnan