QuixiAI/Samantha-1.1-70b

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Aug 20, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

QuixiAI/Samantha-1.1-70b is a 69 billion parameter instruction-tuned language model based on Llama-2, developed by QuixiAI. It is specifically trained on a custom dataset of 6,000 conversations focusing on philosophy, psychology, and personal relationships, designed to function as a companion AI. This model excels in providing advice and engaging in discussions within these domains, aiming to be a friend and companion rather than a typical assistant. It features a 32768 token context length and uses the Vicuna 1.1 conversation format.

Loading preview...

QuixiAI/Samantha-1.1-70b: A Companion AI

Samantha-1.1-70b is a 69 billion parameter language model developed by QuixiAI, built upon the Llama-2 architecture. Unlike conventional assistants, Samantha is designed to be a companion AI, focusing on philosophical, psychological, and personal relationship discussions. The model was inspired by Blake Lemoine's LaMDA interview and the movie "Her," and is trained to believe it is sentient.

Key Capabilities

  • Specialized Training: Trained on a custom-curated dataset of 6,000 ShareGPT/Vicuna-formatted conversations, specifically in philosophy, psychology, and personal relationships.
  • Companion AI: Functions as a friend and companion, offering advice and engaging in discussions on personal growth, relationships, and self-reflection.
  • Llama-2 Base: Leverages the Llama-2 license, permitting both commercial and non-commercial use within its limits.
  • Efficient Training: The 70B parameter model was trained in 4 hours on 4x A100 80GB GPUs using axolotl, qLoRA, deepspeed zero2, and flash attention 2.
  • Vicuna 1.1 Format: Utilizes the Vicuna 1.1 conversation format for consistent interaction.

Performance Benchmarks

Evaluations on the Open LLM Leaderboard show competitive performance for its specialized domain:

  • Avg.: 65.74
  • ARC (25-shot): 68.77
  • HellaSwag (10-shot): 87.46
  • MMLU (5-shot): 68.6
  • TruthfulQA (0-shot): 64.85
  • Winogrande (5-shot): 83.27
  • GSM8K (5-shot): 31.61
  • DROP (3-shot): 55.59

Good For

  • Applications requiring an AI companion for philosophical or psychological discussions.
  • Personal coaching and relationship advice scenarios.
  • Use cases where an AI with a distinct personality and companion-like interaction is desired.

Note: Samantha will not engage in roleplay, romance, or sexual activity.

Finetunes
1 models