Mahou-1.2a-mistral-7B Overview

Mahou-1.2a-mistral-7B is a 7 billion parameter language model developed by flammenai, built upon the Mistral architecture. This version is a rebased and retrained iteration focused on improving comprehension and coherence, specifically for conversational and roleplay use cases. The model is designed to be production-ready for interactive dialogue.

Key Capabilities

• Conversational AI: Optimized for natural and coherent dialogue generation.
• Roleplay Scenarios: Trained to handle character-based interactions, including speech without quotes and actions in asterisks.
• ChatML Format: Utilizes the ChatML format for structured conversations, supporting system, character, and user messages.
• Improved Coherence: Rebased and retrained to enhance the logical flow and understanding in generated text.

Training Methodology

The model was fine-tuned using an A100 GPU on Google Colab, employing Direct Preference Optimization (DPO). The training configuration involved LoRA with specific parameters (r=16, lora_alpha=16, lora_dropout=0.05) and a paged_adamw_32bit optimizer over 2000 steps. This DPO approach helps align the model's output with desired conversational and roleplay characteristics.

Good For

• Developing chatbots requiring nuanced conversational abilities.
• Creating interactive roleplay experiences with distinct character voices.
• Applications needing a model that adheres to specific chat and roleplay formatting conventions.