SurgeFF/AriaV9.2
SurgeFF/AriaV9.2 is a 12 billion parameter multimodal personal assistant model, fine-tuned from Google's Gemma-4-12B-IT. It is optimized for tool calling, memory-aware behavior, and maintaining a stable unprompted identity, with improved mathematical reasoning capabilities. The model features a 32768 token context length and supports vision inputs, making it suitable for complex interactive applications requiring both text and image understanding.
Loading preview...
Aria V9.2: A Multimodal Personal Assistant with Enhanced Math
Aria V9.2 is a 12 billion parameter model, fine-tuned from google/gemma-4-12b-it, designed as a personal assistant. This version significantly improves mathematical reasoning while retaining strong capabilities in tool calling, memory-aware behavior, and a stable unprompted identity. It is an encoder-free multimodal model, meaning vision, audio, and text share weights, ensuring vision capabilities are preserved.
Key Capabilities & Improvements
- Enhanced Mathematical Reasoning: Achieved a notable improvement in math scores, moving from 89% to 91% on a fixed 100-item held-out set, and 87.3% to 92.0% on a fresh 150-item disjoint set. This was accomplished using the STaR (rejection-sampling SFT) method, training on the model's own verified-correct GSM8K solutions.
- Robust Tool Calling: Maintains a perfect 10/10 score for tool calling, supporting specific schemas for five tools (
remember,recall,exec,web_search,send_message). - Memory-Aware Behavior: Shows improved memory behavior, scoring 18/20, indicating it is trained to act correctly around memory systems (though it does not possess intrinsic memory).
- Stable Unprompted Identity: While still scoring 4/8 unprompted, its identity is trained jointly from the first pass, preventing regressions seen in prior attempts to repair identity post-training.
- Multimodal (Vision) Support: Passes vision evaluations, with a 'multimodal floor' in the training data preventing degradation of its vision capabilities.
When to Use This Model
Aria V9.2 is particularly well-suited for applications requiring a personal assistant with:
- Reliable mathematical problem-solving in GSM8K-style contexts.
- Complex interactions involving external tools and structured function calls.
- Context-aware responses that simulate memory, requiring external memory systems.
- Multimodal input processing, especially for tasks involving both text and images.
It's important to note that while math is improved, it's not evaluated on advanced competition problems. The model is tuned for specific conventions and is not a general-purpose assistant release.