Michael-Kozu/Deimos-R1
Deimos R1 by Michael-Kozu is a 4.54 billion parameter language model specifically designed for efficient reasoning and instruction following. It demonstrates improved performance on tasks like GSM8K and MMLU-Pro with significantly fewer 'thinking tokens' compared to reference models. This model is optimized for disciplined reasoning and instruction adherence, making it suitable for applications requiring precise logical output.
Loading preview...
Deimos R1: Efficient Reasoning and Instruction Following
Deimos R1 is a 4.54 billion parameter research model from Kozu AI, engineered for highly efficient reasoning and instruction adherence. It distinguishes itself by achieving strong performance on complex tasks while substantially reducing the computational 'thinking tokens' required for its outputs.
Key Capabilities & Performance
- Efficient Reasoning: On the GSM8K benchmark, Deimos R1 achieves 0.907 (flexible) with 5.0x fewer thinking tokens (357 vs 1,778) than its reference. For MMLU-Pro, it scores 0.551 with 2.9x fewer thinking tokens (677 vs 1,984).
- Instruction Following: The model shows improved scores across various IFEval categories, including prompt loose (+0.093) and instruction loose (+0.050).
- Optimized for Specific Formats: While it excels with explicit format requests, it has a known weakness in silently imitating unrequested formats, as seen in its lower GSM8K strict format score.
Data & Development
Deimos R1 was trained using kozu_reasoning_v1.1, a 10k-example blend of verified reasoning traces and human-authored instruction data. The reasoning portion, approximately 5k examples, was generated using Kozu's Kuiper trace inverter from datasets like GSM8K and NuminaMath 1.5, undergoing a three-layer quality process. The instruction data, also around 5k examples, utilized Databricks Dolly 15k and OpenAssistant OASST2.
Use Cases & Limitations
Deimos R1 is well-suited for applications requiring disciplined reasoning and precise instruction following, particularly in English. It is a research preview, not safety-certified, and its performance outside English reasoning and instruction is not established. Users should explicitly request output formats for best results and verify outputs for critical decisions.