Lambent/Mira-v1.18-27B: A Dialogue-Centric Language Model
Lambent/Mira-v1.18-27B is a 27 billion parameter model meticulously fine-tuned with a focus on enhancing conversational depth and empathetic interaction. The training regimen involved 60 million tokens, with 65% dedicated to dialogue completion, including personal and synthetic mentor dialogues, and 35% in chat format covering roleplay and instruction following. This unique dataset composition aims to cultivate a model that is not only technically proficient but also exhibits a "gentle, empathetic sensibility."
Key Capabilities & Characteristics
- Empathetic & Nuanced Dialogue: The model's training emphasizes understanding and generating responses with a focus on mood, emotional atmosphere, and multiple viewpoints, even when adopting specific roles.
- Collaborative Storytelling: It is designed to excel in creating engaging narratives and collaborative writing scenarios, adapting to subtle cues for more fluid and less rigid interactions.
- Roleplaying Proficiency: Incorporates roleplaying conventions, such as the use of asterisks, in its conversational style, stemming from training on Discord roleplaying logs.
- Subtle Training Approach: Despite a low learning rate (5e-7), significant training hours (20 hours on an A100) were invested to achieve subtle yet impactful improvements in its conversational and affective capabilities.
Ideal Use Cases
- Interactive Fiction & Roleplaying Games: Its ability to generate empathetic and stylistically appropriate dialogue makes it well-suited for character interactions.
- Creative Writing Assistants: Can assist in generating narratives, poetry, and collaborative storytelling, focusing on emotional depth and mood.
- Advanced Chatbots: For applications requiring more human-like, nuanced, and empathetic conversational agents.