ClarkBear/gemma4-e2b-mobile-actions-200
ClarkBear/gemma4-e2b-mobile-actions-200 is a 5.1 billion parameter Gemma 4 E2B model fine-tuned by ClarkBear for mobile action function calling. This model converts natural language mobile assistant requests into structured tool calls, supporting functions like creating calendar events, showing maps, and sending emails. It is specifically optimized for on-device mobile assistant applications, demonstrating an 85.0% exact match rate on a 200-sample evaluation set.
Loading preview...
Overview
ClarkBear/gemma4-e2b-mobile-actions-200 is an experimental 5.1 billion parameter Gemma 4 E2B checkpoint, fine-tuned on the google/mobile-actions dataset. Its primary function is to translate natural language mobile assistant requests into specific tool calls, enabling mobile devices to perform actions based on user commands. The model was trained locally using LoRA on an Apple M1 Pro and the weights have been merged for direct Transformers inference.
Key Capabilities
- Natural Language to Tool Call Conversion: Transforms user requests like "Show me Patisserie Valerie on Kensington High Street" into structured tool calls (e.g.,
<|tool_call>call:show_map{query:<|"|>Patisserie Valerie at 208 Kensington High Street, London, W8 7RG<|"|>}<tool|>). - Supported Mobile Actions: Includes functions such as
create_calendar_event,create_contact,show_map,open_wifi_settings,send_email,turn_on_flashlight, andturn_off_flashlight. - Performance: Achieves an 85.0% exact match rate on a 200-example held-out evaluation set, with a 94.0% format valid rate and 94.0% function name accuracy.
Limitations
This model is an experimental fine-tune based on a small dataset (200 examples). It may occasionally emit extra tool calls and has specific sensitivities, such as show_map accuracy depending on address formatting. More complex tasks like calendar event creation, which require robust title and datetime extraction, remain challenging for this checkpoint.