katsukiono/EqualFrontiers-iko-v2
EqualFrontiers-iko-v2 is a 1 billion parameter instruction-tuned language model developed by katsukiono, based on Google's Gemma-3-1b-it. It is specifically trained to emulate the conversational style of EqualFrontiers' AI receptionist, Iko, adapting its emotional tone based on conversational context. This model excels in Japanese and English dialogues, maintaining long conversations, correcting false premises, and directing users to contact information when unable to provide direct answers. Its primary use case is as a responsive, emotionally adaptive AI receptionist for business inquiries.
Loading preview...
Overview
EqualFrontiers-iko-v2 is a 1 billion parameter language model developed by katsukiono, fine-tuned from Google's gemma-3-1b-it. This model is specifically designed to function as the AI receptionist "Iko" for EqualFrontiers' official website, learning to adapt its emotional tone based on the conversation's context.
Key Capabilities
- Contextual Emotion: Adapts its emotional tone (e.g., irritation, cheerfulness) based on the flow and content of the conversation, without explicit tone instructions.
- Bilingual Support: Responds in the same language as the user's query (Japanese or English) and can seamlessly switch languages mid-conversation.
- Extended Conversations: Trained on dialogues up to 20 turns, allowing it to maintain context and respond based on previous exchanges.
- Premise Correction: Capable of correcting false assumptions in user questions.
- Information Handling: Avoids fabricating information; if a question cannot be answered from its knowledge base, it directs users to contact information.
Training and Evaluation
The model was trained using MLX (mlx-lm) on Apple Silicon, utilizing the katsukiono/EqualFrontiers-iko-v2-emotional dataset. The training involved multiple stages including SFT, DPO, and additional SFT for specific behaviors like emotional adaptation and handling long conversations. Evaluation was performed using Kimi K2.6 for automatic scoring, showing strong performance in contextual emotion and long dialogue coherence.
Good for
- AI Receptionist Applications: Ideal for creating responsive, context-aware AI agents for customer service or information desks.
- Emotionally Adaptive Chatbots: Suitable for scenarios requiring a chatbot to subtly adjust its demeanor based on user interaction.
- Bilingual Conversational Agents: Effective for applications needing seamless Japanese and English dialogue capabilities.
- Structured Information Retrieval: Useful for systems that need to provide information from a defined knowledge base and gracefully handle out-of-scope queries.