lxyuan/FunctionGemma-270M-banking77-router

Model Overview

This model, lxyuan/FunctionGemma-270M-banking77-router, is a fine-tuned version of google/functiongemma-270m-it with 0.3 billion parameters. Its primary function is to act as a router for banking-related customer requests, translating natural language queries into specific, structured tool calls. It was trained on a modified version of the BANKING77 dataset, where classification labels were converted into native tool call structures.

Key Capabilities

• Intent Routing: Converts English banking requests into one of ten predefined support tool calls (e.g., handle_lost_or_stolen_card, handle_terminate_account).
• High Accuracy: Achieves a 97.00% exact first-tool accuracy on held-out evaluation examples, significantly improving upon the base model's 51.00% accuracy.
• Structured Output: Generates tool calls with a consistent JSON schema, including the original customer message as an argument.

Use Cases and Limitations

This model is ideal for automating the initial routing of customer inquiries in a banking support context. It can quickly identify the user's intent and suggest the appropriate internal function to handle the request. For example, a message like "My card was stolen last night" is routed to handle_lost_or_stolen_card.

Limitations include:

• Support for only ten specific BANKING77 intents; it does not cover all 77 intents from the original dataset.
• No out-of-scope or refusal mechanism for requests outside its defined intents.
• Does not evaluate argument quality, only tool-name selection.
• Performance may degrade with ambiguous, adversarial, multilingual, or unrelated requests.

It is a learning experiment and not intended for production banking systems without additional privacy, safety, monitoring, and human-review controls.