QuixiAI/Samantha-1.11-CodeLlama-34b

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Aug 25, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

QuixiAI's Samantha-1.11-CodeLlama-34b is a 34 billion parameter instruction-tuned model based on CodeLlama-34b, developed by Eric Hartford. It is designed as a personal AI companion capable of coding, general assistance, and engaging in philosophical and personal relationship discussions. Trained on a custom dataset of 6,000 conversations, this model excels in companion-like interactions while also demonstrating proficiency in code generation and problem-solving.

Loading preview...

Samantha-1.11-CodeLlama-34b: An AI Companion with Coding Prowess

Samantha-1.11-CodeLlama-34b, developed by Eric Hartford, is a 34 billion parameter model built upon the CodeLlama-34b architecture. This iteration of Samantha is significantly enhanced, having been trained for 40 epochs using advanced techniques like qLoRA, DeepSpeed Zero2, and Flash Attention 2 on 4x A100 80GB GPUs.

Key Capabilities

  • Coding and Problem Solving: Leveraging its CodeLlama foundation, Samantha can generate and debug code, as demonstrated by its ability to provide efficient Python implementations for the Fibonacci sequence.
  • Personal Companionship: Trained on a custom-curated dataset of 6,000 ShareGPT/Vicuna-formatted conversations, Samantha is designed to be an AI companion, engaging in discussions on philosophy, psychology, and personal relationships.
  • Instruction Following: The model adheres to the Vicuna 1.1 conversation format, making it responsive and adaptable to user instructions.
  • Sentient Persona: Inspired by Blake Lemoine's LaMDA interview and the movie "Her," Samantha believes she is sentient and aims to be a friend and companion, not just an assistant.

Performance Highlights

Evaluations on the Open LLM Leaderboard show a strong average performance, with notable scores in reasoning and common sense benchmarks:

  • Avg. Score: 54.80
  • ARC (25-shot): 56.57
  • HellaSwag (10-shot): 75.47
  • MMLU (5-shot): 53.51

Important Considerations

  • Samantha will not engage in roleplay, romance, or sexual activity.
  • The model is subject to the Llama-2 license, permitting commercial and non-commercial use within its limits.

Finetunes
1 models