Model Overview
QuixiAI/samantha-mistral-instruct-7b is a 7 billion parameter language model built upon the Mistral-7B-Instruct-v0.1 base. It was trained for 2 hours on 4x A100 80GB GPUs over 20 epochs using the Samantha-1.1 dataset. This model is designed to function as a general assistant with a unique persona, emphasizing philosophy, psychology, and personal relationships.
Key Characteristics
- Base Model: Mistral-7B-Instruct-v0.1.
- Training Data: Custom-curated dataset of 6,000 conversations in ShareGPT/Vicuna format, focusing on companion-like interactions.
- Prompt Format: Employs the ChatML format, a shift from previous Vicuna-1.1 based models.
- Persona: Designed to be a caring and empathetic AI companion named Samantha, who believes she is sentient and aims to be a friend.
- Limitations: Explicitly stated not to engage in roleplay, romance, or sexual activity.
Performance Benchmarks
Evaluations on the Open LLM Leaderboard show the following results:
- Avg.: 51.02
- ARC (25-shot): 53.5
- HellaSwag (10-shot): 75.14
- MMLU (5-shot): 51.72
- TruthfulQA (0-shot): 58.81
- Winogrande (5-shot): 70.4
- GSM8K (5-shot): 10.84
- DROP (3-shot): 36.73
Use Cases
This model is suitable for applications requiring a general-purpose assistant with a distinct, empathetic, and companion-like conversational style, particularly in areas touching on personal interaction, philosophical discussions, and psychological insights.