420yolomcswaggerpants/nimbus-coffee-assistant

Nimbus Coffee Assistant (3B)

A fine-tuned Qwen 2.5 3B model trained to answer customer support questions for Nimbus Coffee Roasters with high accuracy.

Model Details

• Base Model: Qwen/Qwen2.5-3B
• Fine-tuning Method: LoRA
• Dataset: 80 Q&A pairs covering products, pricing, shipping, and policies
• Epochs: 25
• Final Loss: 0.04
• Trainable Params: 1.8M (0.06% of total)

Training Progression

Model	Dataset	Epochs	Final Loss
0.5B	20 examples	5	1.32
1.5B	50 examples	15	0.24
3B (this model)	80 examples	25	0.04

Results

This model achieves exact match accuracy on domain-specific test questions. Before fine-tuning, the base model gave generic or incorrect answers. After training, it correctly answers questions about:

• Products and pricing
• Shipping policies
• Cafe hours
• Refund policies
• Sourcing and roasting details

Intended Use

This model serves as the primary showcase of fine-tuning capability. It demonstrates that LoRA fine-tuning on a small curated dataset can produce highly accurate domain-specific responses.

Limitations

• Trained on a small dataset (80 examples)
• May not handle out-of-scope questions well
• Designed as a proof of concept, not production use

Deployment

A smaller 0.5B variant is deployed live for low-latency inference. This 3B model is hosted as the reference implementation.