ludis/tsukasa-llama-3-8b-qlora
ludis/tsukasa-llama-3-8b-qlora is an 8 billion parameter instruction-tuned language model based on Meta's Llama 3 8B Instruct architecture. This model has undergone additional training on the Pippa and Limarp datasets, each for two epochs at a 32k context length, enhancing its conversational and roleplay capabilities. It is specifically optimized for generating coherent and engaging responses in interactive chat scenarios, particularly for fictional roleplay.
Loading preview...
Model Overview
ludis/tsukasa-llama-3-8b-qlora is an 8 billion parameter language model built upon the robust Meta Llama 3 8B Instruct foundation. This model has been further refined through an intensive training regimen, undergoing two epochs of training on the Pippa dataset, followed by another two epochs on the Limarp dataset. Both training phases utilized a substantial 32k context length, aiming to enhance the model's ability to handle extended conversations and complex narrative structures.
Key Characteristics
- Base Model: Meta Llama 3 8B Instruct.
- Extended Training: Fine-tuned on Pippa and Limarp datasets for improved performance.
- Context Length: Trained with a 32k context length, supporting longer interactions.
- Instruction Format: Utilizes the Llama 3 instruct format for prompting.
- Stopping Sequence: Employs
<|eot_id|>as the designated end-of-turn token.
Recommended Usage
This model is particularly well-suited for fictional roleplay chat and generating dynamic, character-driven dialogues. The README provides specific generation settings and prompting examples, including an agnostic prompt structure, to help users achieve optimal results in roleplay scenarios. Users are advised to experiment with the suggested sampler settings (Temperature, Min-P, Presence Penalty, Frequency Penalty, or Mirostat) to fine-tune output coherence and creativity, especially to mitigate potential repetition issues.