SurgeFF/AriaV9.1

Aria V9.1: A Specialized Personal Assistant

Aria V9.1 is a 12 billion parameter model, fine-tuned from google/gemma-4-12b-it, designed as a personal assistant with enhanced capabilities in tool calling, memory-aware behavior, and a stable, unprompted identity. It maintains the base model's strong math and vision capabilities.

Key Enhancements in V9.1

This version primarily addresses and improves the model's unprompted identity, a known limitation in its predecessor (V8). While V8 struggled to identify itself without a system prompt, V9.1 significantly improves this by integrating identity training directly into the first pass, avoiding performance degradation seen in repair attempts. Evaluation shows unprompted identity improved from ~0/8 to 4/8, with other capabilities like math (89/100 on GSM8K) and tool calling (10/10) remaining consistent.

Training and Architecture

Aria V9.1 was trained on a single RTX 3090 using a legally clean corpus generated by open-weight teacher models. It utilizes a single-stage data mix including specific datasets for tool-calling, memory-aware behavior, and identity. As gemma-4 is an encoder-free multimodal model, a deliberate multimodal floor was included in training to prevent vision degradation. The fine-tuning process involved adjusting lm_head and embed_tokens to optimize output-token behavior.

Good for

• Personal assistant applications requiring consistent identity and memory management.
• Tool-integrated systems using specific tool call conventions (e.g., recall, exec, web_search).
• Multimodal tasks where vision capabilities are important, though vision is verified, not optimized for benchmarks.
• Mathematical reasoning on GSM8K-style problems.