ayoubkirouane/Sophea-Nemo-3.5-Lightning-v1
Sophea-Nemo-3.5-Lightning-v1 is a 30B parameter fine-tuned version of the Nemotron-3.5-Lightning-30B-A3B (Mamba/MoE hybrid) model, developed by Kiefer SA (Sophea AI Lab, Athens). It specializes in Greek and English reasoning, producing thinking traces in the question's language. This model demonstrates improved Greek NLU performance (+1.7%) with a minor English NLU cost (-1.1%), making it suitable for bilingual reasoning applications requiring strong Greek language capabilities.
Loading preview...
Sophea-Nemo-3.5-Lightning-v1: Bilingual Reasoning with Nemotron-3.5-Lightning
Sophea-Nemo-3.5-Lightning-v1 is a 30B parameter model developed by Kiefer SA (Sophea AI Lab, Athens), fine-tuned from the nvidia/Nemotron-3.5-Lightning-30B-A3B base. This model is designed for Greek and English reasoning, generating thinking traces within <think> blocks that follow the language of the input question. It was released alongside the paper "Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See" (arXiv:2608.17744).
Key Capabilities & Features
- Bilingual Reasoning: Excels in generating reasoning traces for both Greek and English questions, with 98.1% Greek-trace fidelity for Greek questions and 100% English traces for English questions.
- Strong Greek NLU: Achieves a +1.7% gain in Greek NLU macro score compared to its base model, demonstrating improved performance in a low-resource language.
- Minimal English Forgetting: Experiences only a -1.1% decrease in English NLU macro score, indicating good retention of general English capabilities.
- Controlled Reasoning Language: While it does not obey instructions to force English traces on Greek questions (0.0% override), it successfully switches to Greek traces for English questions 92.7% of the time.
- Hybrid Architecture: Inherits the Nemotron-3.5-Lightning-30B-A3B's next-gen Mamba/MoE hybrid architecture (31.6B total / 3.58B active parameters).
- Speculative Decoding (MTP): Supports multi-token-prediction (MTP) for enhanced throughput, utilizing the base model's
mtp.*tensors.
Intended Use Cases
- Deployments requiring robust Greek and English reasoning capabilities.
- Applications where the reasoning trace should naturally follow the input question's language.
- Scenarios benefiting from the Nemotron-line hybrid architecture with a mild logic fallback (5.3%).
Limitations
- Not evaluated for safety-critical, legal, or medical use.
- Cannot be forced to reason in English for Greek questions.