ayoubkirouane/Sophea-Nemo-3-Nano-v1
Sophea-Nemo-3-Nano-v1 by Kiefer SA (Sophea AI Lab, Athens) is a 30 billion parameter Mamba/MoE hybrid model, fine-tuned from nvidia/NemotronH-30B-A3B. It specializes in Greek and English reasoning, demonstrating high fidelity in generating reasoning traces in the question's language with 0.0% answer-channel leak. This model is optimized for applications requiring robust reasoning in both Greek and English, particularly where maintaining answer integrity is critical.
Loading preview...
Overview
Sophea-Nemo-3-Nano-v1 is a 30 billion parameter language model developed by Kiefer SA (Sophea AI Lab, Athens), fine-tuned from the nvidia/NemotronH-30B-A3B base model, which features a Mamba/MoE hybrid architecture. This model is specifically designed for reasoning tasks in both Greek and English, with a focus on maintaining the integrity of reasoning traces within the question's language.
Key Capabilities
- Bilingual Reasoning: Excels at generating reasoning traces in Greek and English, with 97.4% Greek-trace fidelity for Greek questions and 100% English traces for English questions.
- Answer-Channel Integrity: Achieves 0.0% answer-channel leak, ensuring that reasoning steps do not inadvertently appear in the final answer.
- Minimal Forgetting: The fine-tuning process resulted in a gain of +3.8 in Greek NLU macro performance on the Titan-1 suite, with only a minor -2.0 English NLU macro change, indicating robust retention of general language abilities.
- Efficient Tracing: Median trace length is 638 tokens, comparable to the base model, and shows significant reduction in generation-cap truncation and measured loops in traces.
Good For
- Applications requiring reliable Greek and English reasoning where the reasoning trace must strictly follow the question's language.
- Deployments that benefit from the base model's general Greek language capabilities, as the fine-tune enhances this aspect.
- Use cases where preventing reasoning steps from leaking into the direct answer is paramount.