Model Overview

DavidAU/Qwen3-Grand-Horror-DARK-Child-1.7B is a 2 billion parameter model, fine-tuned by DavidAU, specifically designed for generating horror content. It leverages an in-house horror dataset and Unsloth for tuning, resulting in outputs that consistently feature elements of horror, madness, swearing, and gore.

Key Characteristics

• Horror-Centric Output: The model's strong tuning ensures that almost all generations, including stories, roleplay adventures, fiction, and even general replies, will have a horror tinge, regardless of the prompt's original intent.
• Context Length: Supports a substantial 40,960 token context window, with a recommended minimum of 4,000 tokens for long generations.
• Output Intensity: Capable of generating mild to strong levels of horror.
• Parameter Recommendations: Optimal performance is achieved with a repetition penalty of 1.1 (or 1.15) and a temperature range of 0.4 to 1.2, with higher temperatures (1.5+) also yielding excellent results.

Usage Considerations

• Repetition: Due to its size, the model may occasionally repeat paragraphs or words, particularly with lower quantization levels (5 bits or less).
• Smoothing Factor: For smoother operation in applications like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern, setting a "Smoothing_factor" of 1.5 is recommended.
• GGUF Compatibility: When using GGUFs in text-generation-webui, the llama_HF loader is required, which necessitates downloading config files from the model's source version.