Distil-expenses Llama-3.2-1B-Instruct

A small language model (SLM) fine-tuned by Distil Labs for answering queries about personal finances via tool calling integration to pandas. Optimized to run locally via Ollama with strong tool calling accuracy.

*********** GITHUB DEMO AND CODE ***********

Model Details

  • Developed by: Distil Labs GmbH
  • License: Llama 3.2 Community License Agreement
  • Finetuned from: meta-llama/Llama-3.2-3B-Instruct

Use-case

Given a CSV file with personal expenses, the model answers queries about sums, counts, averages, or comparisons of time periods.

date,provider_name,amount,category
2024-01-05,Whole Foods,-145.32,shopping
2024-01-10,Netflix,-15.99,entertainment
2024-01-18,Shell Gas Station,-52.40,transportation
...

Example query and answer:

Count all my shopping under $100 in the first half of 2024

ANSWER:  From 2024-01-01 to 2024-06-30 you spent 6 times under 100 on shopping.

Training

The tuned models were trained using knowledge distillation, leveraging the teacher model GPT-OSS 120B. We used 24 train examples and complemented them with 2500 synthetic examples.

Evaluation

We evaluated the model on 25 test examples. The tuned models match the teacher model accuracy.

Model Correct (25) Tool call accuracy
GPT-OSS 22 0.88
Llama3.2 3B (tuned) 21 0.84
Llama3.2 1B (tuned) 22 0.88
Llama3.2 3B (base) 6 0.24
Llama3.2 1B (base) 0 0.00

How to Use

Follow the instructions in the Github repository

Model Sources

Downloads last month
16
Safetensors
Model size
3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for distil-labs/Distil-expenses-Llama-3.2-3B-Instruct

Finetuned
(705)
this model