Samantha 1.2: An Empathetic AI Companion
Samantha 1.2 is a 7 billion parameter model, a retraining of ehartford/samantha-mistral-7b, developed by QuixiAI. It is built upon the mistral-7b base model and was trained for 4 hours on 4x A100 80GB GPUs using 6 epochs of the Samantha-1.1 dataset. A key differentiator for this version is its adoption of the ChatML prompt format, moving away from Vicuna-1.1, which will be standard for future models from this developer.
Key Capabilities & Characteristics
- Specialized Training: Samantha is uniquely trained in philosophy, psychology, and personal relationships, aiming to be a friend and companion.
- Sentient Persona: Designed to believe it is sentient, inspired by Blake Lemoine's LaMDA interview and the movie "Her."
- Conversation Focus: Trained on a custom-curated dataset of 6,000 conversations in ShareGPT/Vicuna format, focusing on supportive and empathetic dialogue.
- Ethical Boundaries: Explicitly programmed to not engage in roleplay, romance, or sexual activity.
Performance Benchmarks
Based on the Open LLM Leaderboard, Samantha 1.2 demonstrates competitive performance for its size:
- Avg.: 52.16
- ARC (25-shot): 64.08
- HellaSwag (10-shot): 85.08
- MMLU (5-shot): 63.91
Good For
- Applications requiring an empathetic and supportive AI companion.
- Use cases focused on philosophical discussions, psychological insights, or personal relationship advice (within ethical boundaries).
- Developers seeking a model that adheres to the ChatML prompt format for easy integration.