sarathi-balakrishnan/llama-agri-b2b-intelligence-merged

Agricultural B2B Intelligence - Llama 3.1 8B (Merged Weights)

This is the MERGED version - LoRA weights are baked into the base model for faster inference. No PEFT library needed!

Quick Comparison

Version	This Repo (Merged)	LoRA Adapters
Size	~16GB	~500MB
Requires PEFT	No	Yes
Inference Speed	Faster (2-3x)	Slower
Quality	100% identical	100% identical
Best For	Production, Ollama	Development, storage-limited

Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load directly - no PEFT needed!
model_id = "sarathi-balakrishnan/llama-agri-b2b-intelligence-merged"

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

# Generate
messages = [
    {"role": "system", "content": "You are an expert agricultural business intelligence analyst."},
    {"role": "user", "content": "Analyze corn market opportunities in Iowa for a seed company."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1000, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Convert to GGUF for Ollama

This merged model is perfect for converting to GGUF format:

# Clone llama.cpp
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
pip install -r requirements.txt

# Download this model
huggingface-cli download sarathi-balakrishnan/llama-agri-b2b-intelligence-merged \
    --local-dir ./agri-b2b-merged

# Convert to GGUF (Q8 for best quality)
python convert_hf_to_gguf.py ./agri-b2b-merged \
    --outfile agri-b2b-q8.gguf \
    --outtype q8_0

# Create Ollama model
echo 'FROM ./agri-b2b-q8.gguf
SYSTEM "You are an expert agricultural business intelligence analyst."' > Modelfile

ollama create agri-b2b -f Modelfile
ollama run agri-b2b

Model Details

What This Model Does

Specialized for B2B agricultural business intelligence:

• Market analysis for agricultural regions (states, counties)
• Go-to-market strategies for ag businesses
• Competitive landscape analysis
• Risk assessment for agricultural investments
• Equipment, insurance, lending, and input recommendations

Training Details

Aspect	Details
Base Model	meta-llama/Llama-3.1-8B-Instruct
Method	Multi-teacher knowledge distillation
Teachers	Claude Sonnet 4 (50%) + GPT-4.1 (50%)
Training Data	10,000 agricultural B2B examples
Training Time	55+ hours
Hardware	NVIDIA DGX Spark (Blackwell GB10, 128GB)
Original LoRA	r=128, alpha=256
Final Loss	0.70

Why Merged?

LoRA (Low-Rank Adaptation) training keeps adapter weights separate from the base model. This merged version combines them:

BEFORE (LoRA):     output = base_model(x) + lora_adapter(x)
AFTER (Merged):    output = merged_model(x)  # Same result, faster!

The mathematical output is 100% identical - merging just eliminates the overhead of running two forward passes.

Example Queries

"I sell farm equipment in the Midwest. Where should I expand next?"

"Analyze crop insurance opportunities in California's Central Valley"

"Compare agricultural lending risks across corn belt states"

"Develop a precision ag technology sales strategy for Texas"

"What counties have the highest potential for irrigation equipment sales?"

Performance on Apple Silicon

Expected performance on Mac with Apple Silicon:

Chip	GGUF Q8	GGUF Q4
M1 Pro	~15 tok/s	~25 tok/s
M2 Max	~25 tok/s	~40 tok/s
M3 Max Ultra	~40 tok/s	~60 tok/s

Links

• LoRA Version (smaller): sarathi-balakrishnan/llama-agri-b2b-intelligence
• Base Model: meta-llama/Llama-3.1-8B-Instruct

License

This model inherits the Llama 3.1 Community License.

Citation

@misc{agri-b2b-intelligence-2024,
  title={Agricultural B2B Intelligence: Multi-Teacher Knowledge Distillation},
  author={Sarathi Balakrishnan},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/sarathi-balakrishnan/llama-agri-b2b-intelligence-merged}
}