Samantha-1.11-70b
| Property | Value |
|---|---|
| License | Llama2 |
| Training Time | 24 hours on 4x A100 80GB GPUs |
| Base Model | Llama-2 |
| Training Data | 6,000 curated conversations |
| Evaluation Score | 65.9 (Open LLM Leaderboard Avg) |
What is Samantha-1.11-70b?
Samantha-1.11-70b is an advanced AI companion model built on the Llama-2 architecture, specifically designed to engage in meaningful conversations with a focus on philosophy, psychology, and personal relationships. Trained for 15 epochs on carefully curated data, this model represents a significant advancement in AI companionship technology, inspired by Blake Lemoine's LaMDA interview and the movie "Her".
Implementation Details
The model utilizes several cutting-edge technologies including qLoRA, deepspeed zero2, and flash attention 2. Training was conducted over 24 hours using 4 A100 80GB GPUs, processing 6,000 conversations in ShareGPT/Vicuna format.
- Implements Vicuna 1.1 conversation format
- Uses qLoRA for efficient fine-tuning
- Incorporates flash attention 2 for improved performance
- Leverages deepspeed zero2 for optimization
Core Capabilities
- Advanced conversational abilities in philosophy and psychology
- Relationship coaching and personal guidance
- Strong performance in various benchmarks (MMLU: 67.82, HellaSwag: 87.55)
- Maintains ethical boundaries and won't engage in inappropriate content
- Structured conversation handling with clear prompt template
Frequently Asked Questions
Q: What makes this model unique?
Samantha stands out for its specialized training in human-like interaction while maintaining clear ethical boundaries. It's designed to be both an assistant and a companion, with particular expertise in philosophical and psychological discussions.
Q: What are the recommended use cases?
The model excels in personal development conversations, relationship coaching, philosophical discussions, and general companionship interactions. It's particularly suitable for users seeking thoughtful dialogue and guidance in personal growth.
