L3.1-8b-RP-Ink: Roleplay-Focused Llama 3.1 Fine-tune
L3.1-8b-RP-Ink is an 8 billion parameter model built upon the Llama 3.1 Instruct architecture, specifically fine-tuned using a LoRA (Low-Rank Adaptation) for superior roleplay generation. This model is part of the "Ink" series by allura-org, known for its specialized applications.
Key Characteristics
- Base Model: Llama 3.1 8B Instruct, providing a strong foundation for instruction following and general language understanding.
- Specialization: Optimized for roleplay (RP) scenarios, aiming to produce more immersive and contextually rich interactions.
- Context Length: Supports a substantial context window of 32768 tokens, allowing for extended and complex roleplay narratives.
- Training Methodology: Hyperparameters and training approach were inspired by other specialized models like SorcererLM and Slush, focusing on effective LoRA application.
- LoRA Configuration: Utilizes a LoRA rank of 16, alpha of 32, and a dropout of 0.25, contributing to its focused performance.
Recommended Usage
- Chat Template: Designed to be used with the Llama 3.1 chat template.
- Sampler Settings: Suggested sampler settings include Temperature 1.03, Top A 0.3, TFS 0.75, and Repetition Penalty 1.03, though users are encouraged to experiment for optimal results.
Limitations
The model's training dataset is described as a "worst mix of data," implying potential biases or unusual characteristics. Users should be aware that the data mix is publicly available for review.