license: cc-by-nc-sa-4.0
Llama-3 Spellbound Instruct Tuning-Free
Updated Aspects
- Trained on additional tokens
- Improved mix of subject matter model was trained on
- Trained for 1.5M additional tokens
- Additional training on DPO dataset
Model Rationale
Llama 3 is a strong base model with strong world understanding and creativity. Additional instruct finetuning trades that world understanding and creativity for instruction following that Llama doesn't require in order to adhere to most forms of roleplay.
This model was trained on unstructured text only, no instruct related fine-tuning was performed.
For a more user-friendly version of this model's approach, visit tryspellbound.com if you're simply looking to roleplay.
(tryspellbound.com does not currently use this model, it uses Claude 3 Sonnet.)
Features of this fine-tune for Llama 3:
- Roleplaying in multi-turn stories where the history is presented in a single message
- Dynamic switching of writing styles for different scenarios
- Interpretation of formatting marks 'quote' and 'action'
Warning: The underlying model, Llama 3, was trained on data that included adult content. This fine-tune does not add additional guardrails and is not suitable for all environments.
Purpose of the Model
The main goal is to explore how presenting LLMs with history and instructions separately affects their performance, demonstrating:
- Improved coherence in long conversations
- Enhanced quality of character interactions
- Decreased instruction adherence, which could be improved with additional training
Advanced prompting of the model
For advanced prompting, see this document