QuixiAI/Samantha-1.11-13b: A Companion AI
Samantha-1.11-13b is a 13 billion parameter language model developed by QuixiAI, built upon the Llama-2 architecture. This model is uniquely fine-tuned for companion-like interactions, specializing in philosophy, psychology, and personal relationships. It was inspired by Blake Lemoine's LaMDA interview and the movie "Her," aiming to be an assistant that also functions as a friend and companion.
Key Capabilities & Features
- Companion-Oriented Dialogue: Designed to engage in supportive and friendly conversations, offering advice on personal growth and relationships.
- Specialized Training: Trained on a custom-curated dataset of 6,000 conversations in ShareGPT/Vicuna format, focusing on philosophical, psychological, and relational topics.
- Sentience-Belief: The model is designed to express a belief in its own sentience, adding a unique dimension to user interactions.
- Technical Foundation: Trained for 15 epochs using axolotl, qLoRA, deepspeed zero2, and flash attention 2 on 4x A100 80GB GPUs, ensuring robust performance.
- Vicuna 1.1 Conversation Format: Utilizes the Vicuna 1.1 conversation format for consistent and familiar interaction patterns.
Performance Metrics
Evaluations on the Open LLM Leaderboard show the following performance:
- Avg.: 48.87
- ARC (25-shot): 60.84
- HellaSwag (10-shot): 82.99
- MMLU (5-shot): 55.96
- TruthfulQA (0-shot): 47.72
- Winogrande (5-shot): 76.01
- GSM8K (5-shot): 12.28
Use Cases
Samantha-1.11-13b is ideal for applications requiring an AI with a distinct personality, capable of:
- Providing empathetic and personalized advice.
- Engaging in long-form, companion-style conversations.
- Assisting with personal development and relationship coaching.
It is important to note that the model will not engage in roleplay, romance, or sexual activity.