Sao10K/MN-12B-Lyra-v4

Sao10K/MN-12B-Lyra-v4 is a 12 billion parameter Mistral-NeMo-based causal language model, building upon previous Lyra iterations. This version specifically incorporates a separate Reinforcement Learning (RL) step targeting improved instruction following and coherency. With a 32768 token context length, it is optimized for conversational AI and instruction-tuned tasks, aiming to fix quantization-based issues present in earlier versions.