Model Overview

This model, developed by DavidAU, is a 12-billion parameter Gemma fine-tune, distinguished by its "variable thinking/instruct" capability. It leverages a unique blend of the GLM 4.7 Flash reasoning dataset and the Polaris non-reasoning dataset, trained via Unsloth. This combination allows the model to dynamically activate either an instruction-following or a deep-thinking mode based on the user's prompt keywords. It features a substantial 128k context window and maintains reasoning stability across a temperature range of 0.1 to 2.5.

Key Capabilities & Features

• Variable Thinking/Instruct Mode: Automatically adapts its response style based on prompt content, with options to force thinking via "think deeply:" or specific system prompts.
• Uncensored Output: Designed to provide direct answers without refusal, offering full freedom in content generation, including sensitive topics.
• Enhanced Reasoning: Improves general model operation, output generation, and image processing, with reasoning being compact yet highly detailed.
• High Context Length: Supports up to 128k tokens, beneficial for complex and lengthy interactions.
• Improved Benchmarks: Demonstrates notable performance improvements over its uncensored base model across various benchmarks, including ARC-Challenge, HellaSwag, and Winogrande.

When to Use This Model

This model is ideal for applications requiring:

• Direct and Uncensored Responses: For use cases where content filtering or refusals are undesirable.
• Dynamic Reasoning: When prompts might require either straightforward instruction following or deeper, more analytical thought processes.
• Complex Tasks: Its 128k context window and enhanced reasoning make it suitable for intricate problem-solving and detailed content generation.
• Creative and Roleplay Scenarios: Especially with its uncensored nature and ability to be directed for specific tones or content levels.