Qwen3 BankAssist

Qwen3 BankAssist is a banking intent classification model based on Qwen3-0.6B.

It classifies customer banking messages into one of 77 intent categories from the Banking77 dataset.

Run

”`bash ollama run waseemghanem98/qwen3-bankassist Example

Input:

My card still hasn’t arrived.

Output:

card_arrival Model Details Base model: Qwen/Qwen3-0.6B Task: Banking intent classification Dataset: Banking77 Intent classes: 77 Parameters: 596M Format: GGUF Precision: F16 Runtime: Ollama Evaluation

The model was evaluated on the full Banking77 test split containing 3,076 unseen examples across all 77 intent categories.

Model Accuracy Invalid predictions Qwen3-0.6B 22.07% 335 Qwen3 BankAssist 47.79% 160

The fine-tuned model improved accuracy by 25.72 percentage points and reduced invalid predictions from 335 to 160.

Evaluation used the same prompt, tokenizer, deterministic generation settings, and exact intent-label matching for both models.

Model Pipeline Banking77 ↓ Qwen3-0.6B ↓ LoRA fine-tuning ↓ Merged model ↓ GGUF ↓ Ollama Usage

The model is intended for experiments involving banking intent classification, NLP workflows, and local LLM inference.

Example messages include:

Where is my card? I forgot my PIN. Why was my transfer declined?

The expected output is a single Banking77 intent label.

Related Models

Hugging Face standalone model:

WaseemGh98/bankassist-qwen3

Hugging Face LoRA adapter:

WaseemGh98/waseem-bankassist-qwen3-lora

Limitations

The model is experimental and is not intended for production banking decisions or customer-facing financial advice.

Its measured accuracy on Banking77 is 47.79%, so predictions may be incorrect or fall outside the expected intent labels.

Attribution

Base model: Qwen/Qwen3-0.6B Base model license: Apache 2.0

Dataset: Banking77 Original dataset: PolyAI/banking77 Dataset license: CC BY 4.0