Broken-Tutu-24B-Transgression-v2.0: Enhanced Narrative and Character Integrity
Broken-Tutu-24B-Transgression-v2.0 is a 24 billion parameter model developed by ReadyArt, sleepdeprived3, Artus, gecfdo, and mradermacher, fine-tuned from mistralai/Mistral-Small-24B-Instruct-2501. This iteration focuses on improving coherence and character authenticity while significantly reducing explicit content compared to its predecessors. It leverages "Transgression" techniques for superior performance in creative writing and roleplay.
Key Capabilities & Features
- Expanded 43M Token Dataset: The first ReadyArt model to incorporate multi-turn conversational data, enhancing dialogue and interaction quality.
- 100% Unslopped Dataset: Utilizes new techniques for dataset generation, ensuring high data quality with 0% "slop."
- Enhanced Character Integrity: Maintains character authenticity and core traits (e.g., wholesome characters remain wholesome, yanderes remain intense) while generating less explicit content.
- Anti-Impersonation Guards: Prevents the model from speaking or acting on behalf of the user, ensuring a clear distinction in roleplay.
- Optimized Training: Rebuilt from the ground up with optimized training settings (QLoRA with DeepSpeed Zero3, 5120 sequence length, 2e-6 learning rate) for superior performance.
- Reduced Repetition and Hallucination: Offers improved narrative flow and factual consistency compared to previous versions.
- Superior Instruction Following: Excels at understanding and executing complex prompts, adapting to subtle nuances.
- Enhanced Image Understanding: Features capabilities for multimodal interactions, suggesting potential for image-based prompts.
Ideal Use Cases
This model is particularly well-suited for:
- Roleplay and Storytelling: Designed for engaging in long-form, multi-character scenarios with strong narrative coherence.
- Creative Writing: Generating detailed and consistent narratives while maintaining specific character traits.
- Character-Driven Interactions: When maintaining strict character authenticity and avoiding model impersonation is crucial.
- Applications Requiring Reduced Explicit Content: For users seeking a more balanced approach to content appropriateness in creative generation.