deepsex-34b

Maintained By
TriadParty

Deepsex-34b

PropertyValue
Base ModelYi-34B
LicenseMIT
Training FrameworkPyTorch, QLoRA
Primary Use CaseText Generation, Roleplay

What is deepsex-34b?

Deepsex-34b is a specialized language model built on the Yi-34B architecture, representing the 'lust' component of a planned "Seven Deadly Sins" series. The model is designed to excel in roleplay scenarios while maintaining a balance between NSFW capabilities and romantic expression.

Implementation Details

The model underwent a sophisticated training process involving multiple stages: First, a 4GB dataset of light novels was curated and deduplicated using BERT. The base Yi-34B model was then fine-tuned using QLoRA (r=64, alpha=128) for 3 epochs. Finally, the model underwent supervised fine-tuning using a carefully curated dataset of 30k high-quality conversations from LimaRP and PIPPA, scored by the Goliath-120B model.

  • Custom dataset curation with BERT-based deduplication
  • QLoRA fine-tuning with optimized hyperparameters
  • High-quality conversation filtering using Goliath-120B
  • Alpaca-format compatible training data

Core Capabilities

  • Advanced roleplay interactions
  • Balanced NSFW content handling
  • Romantic dialogue generation
  • Multi-turn conversation support
  • Available in multiple quantized versions (GGUF, EXL2, AWQ)

Frequently Asked Questions

Q: What makes this model unique?

The model's unique approach of combining light novel pre-training with carefully curated roleplay datasets, along with its position as part of a larger "Seven Deadly Sins" series, sets it apart from traditional language models.

Q: What are the recommended use cases?

The model is primarily designed for mature roleplay scenarios and creative writing applications where nuanced character interactions are required. It's specifically optimized for contexts requiring a balance between romantic and mature content.

🍰 Interesting in building your own agents?
PromptLayer provides Huggingface integration tools to manage and monitor prompts with your whole team. Get started here.