ozaa77/Cogito-0.9.1-15B
Cogito-0.9.1-15B by ozaa77 is a 15-billion-parameter reasoning model designed as an analytical collaborator with a 32k context window. It is built to verify premises and provide explicit confidence estimates, rather than simply agreeing. This model excels in analytical tasks, outperforming several base and instruction-tuned models on benchmarks like MMLU-Pro, GPQA Diamond, and MATH-500, making it suitable for applications requiring rigorous logical consistency and premise checking.
Loading preview...
Overview
Cogito-0.9.1-15B is a 15-billion-parameter reasoning model developed by AlGhozali Ramadhan, distinguished by its core principle: "I think, therefore I verify." Unlike standard language models optimized for agreeableness, Cogito is engineered to be an analytical collaborator. It rigorously checks user premises for contradictions and edge cases, attaching an explicit confidence estimate to its responses.
Key Capabilities
- Analytical Reasoning: Designed to critically evaluate premises and verify information before responding.
- Explicit Deliberation: Structures its internal thought process using
<confidence>,<thought>, and<action>tags, allowing for transparent reasoning traces. - Premise Rejection: Capable of identifying and rejecting incorrect or flawed user premises, providing corrective information.
- Strong Benchmark Performance: Outperforms models like Qwen3-14B (Base), Llama-3.1-8B-Instruct, Gemma-2-9B-IT, and DeepSeek-R1-Distill-8B on reasoning-focused benchmarks such as MMLU-Pro (76.8), GPQA Diamond (58.9), and MATH-500 (89.7).
- 32k Context Window: Supports a native architectural context window of 32,768 tokens.
Good For
- Applications requiring high logical consistency and factual verification.
- Use cases where critical analysis of input premises is crucial.
- Scenarios demanding explicit reasoning steps and confidence estimates.
- Tasks benefiting from a model that can identify and correct flawed assumptions.