Synthia S1 27B - AI Model for Creative Writing - Install and Test Locally



AI Summary

Cynthia Model Overview

  • Introduction: Cynthia S1 is a creative writing model fine-tuned from the Gemma 3 architecture.
  • Capabilities:
    • Excels in logical reasoning and deep contextual understanding.
    • Supports multimodal input (text and images).
    • Large 128k token context window for complex analysis.

Installation Guide:

  1. Setup: Using a VM Ubuntu with Nvidia H100 GPU (80 GB VRAM).
  2. Create Virtual Environment.
  3. Install Prerequisites:
    • Use VLM for inference.
    • Install Uber Booker’s text generation web UI.
  4. Download Model:
    • Log into Hugging Face, create a directory named ‘Cynthia’, and download the model (12 shards).
  5. Launching:
    • Access UI at localhost, load the model onto the GPU.

Configuration Settings:

  • Hyperparameters:
    • Temperature: 1 (controls randomness).
    • Top K: 64 (limits token choices).
    • Top P: 95 (nucleus sampling).
    • Min P: 0.0 (no minimum cutoff).
    • Repetition Penalty: 1.3 (reduces repeated tokens).
  • Chat Interface: Engage in creative writing with various prompts, including:
    1. Story of the world’s oldest tree finding an impossible blossom.
    2. Pandemic-themed laughter contagion.
    3. Society where memories are currency.

Writing Quality: Prompts showcase the model’s creative capabilities, producing imaginative and original stories with room for improvement in some narrative elements. Overall, it effectively captures themes and engages the reader.