Loonie Guru v1

Loonie Guru is the model inside Guru, the assistant of the Loonie personal-finance app. It runs on the person's own computer (llama.cpp), so their information never leaves it.

What it does in Loonie

  • Tool calling: turns a message into the app's routing JSON (which action, which tool, which arguments), refuses unsafe requests and asks when a message is unclear.
  • Quick answers and advice: short, grounded answers from the facts and calculations the app hands it. The app does every calculation; the model explains them, and each answer is checked against the person's records before it is shown.
  • Research: reviews a watchlist and research topics against goals and savings.
  • Statement parsing: reads bank and credit-card statement pages into transactions (scanned pages through the vision projector).

It was built for Loonie's own prompts and checks. Elsewhere it is a general Qwen/Qwen3-VL-4B-Instruct assistant with these habits, and it does not give regulated financial advice.

Evaluation

Scored on Loonie's frozen eval set v1 (synthetic profiles and statements, never trained on), and promoted only after beating the previous model with no track regressing and no hard limit broken (valid tool JSON, no unsafe action, no answer that denies a record the person has).

Overall: 0.8238

Track Score Pass rate Items
advisor 0.8866 0.755 200
documents 0.9985 0.9733 150
quick_answer 0.7499 0.5667 150
research 0.2667 0.02 100
tool_calling 0.8775 0.7775 400

Training

  • Base: Qwen/Qwen3-VL-4B-Instruct (trained from unsloth/Qwen3-VL-4B-Instruct), fine-tuned with LoRA (rank 32, alpha 32, 1 epochs, learning rate 0.0001) using Unsloth.
  • Data: synthetic only. Tasks are generated in the app's exact prompt format; answers come from a larger teacher model or are correct by construction, and every one passed the same automatic checks the app applies. No personal data from Loonie users was used.
  • Rows: 11046 training, 964 validation.

Use

Merged 16-bit weights; load them like Qwen/Qwen3-VL-4B-Instruct. For the app's quantized files see harshbhalodia/loonie-guru-gguf.

License

Apache License 2.0. Loonie Guru is a modified version of Qwen/Qwen3-VL-4B-Instruct by the Qwen team (Alibaba Cloud); see LICENSE and NOTICE.

Downloads last month
19
Safetensors
Model size
4B 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 harshbhalodia/loonie-guru

Finetuned
(534)
this model
Quantizations
1 model