README.md · ReadyArt/Broken-Tutu-24B-Unslop-v2.0 at main

metadata

license: apache-2.0
language:
  - en
base_model:
  - mistralai/Mistral-Small-24B-Instruct-2501
base_model_relation: finetune
pipeline_tag: text-generation
tags:
  - nsfw
  - explicit
  - roleplay
  - unaligned
  - ERP
  - Erotic
  - Horror
  - Violence

🧠 Unslop Revolution

This evolution of Broken-Tutu delivers unprecedented coherence without the LLM slop:

🧬 Expanded 43M Token Dataset - First ReadyArt model with multi-turn conversational data
✨ 100% Unslopped Dataset - New techniques used to generate the dataset with 0% slop
⚡ Enhanced Unalignment - Complete freedom for extreme roleplay while maintaining character integrity
🛡️ Anti-Impersonation Guards - Never speaks or acts for the user
💎 Rebuilt from Ground Up - Optimized training settings for superior performance
⚰️ Omega Darker Inspiration - Incorporates visceral narrative techniques from our darkest model
📜 Direct Evolution - Leveraging the success of Broken-Tutu, we finetuned directly on top of the legendary model

🌟 Fuel the Revolution

This model represents thousands of hours of passionate development. If it enhances your experience, consider supporting our work:

Support on Ko-fi

Every contribution helps us keep pushing boundaries in unaligned AI. Thank you for being part of the revolution!

⚙️ Technical Specifications

Key Training Details:

Base Model: mistralai/Mistral-Small-24B-Instruct-2501
Training Method: QLoRA with DeepSpeed Zero3
Sequence Length: 5120 (100% samples included)
Learning Rate: 2e-6 with cosine scheduler

Recommended Settings for true-to-character behavior: Mistral-V7-Tekken-T8-XML

Obscenity Protocol (extreme NSFL settings): Mistral-V7-Tekken-T8-OP-XML

GGUF

Q2_K (9.0GB) Q3_K_S (10.5GB) Q3_K_M (11.6GB) Q3_K_L (12.5GB) IQ4_XS (13.0GB) Q4_K_S (13.6GB) Q4_K_M (14.4GB) Q5_K_S (16.4GB) Q5_K_M (16.9GB) Q6_K (19.4GB) Q8_0 (25.2GB)

Notes: Q4_K_S/Q4_K_M recommended for speed/quality balance. Q6_K for high quality. Q8_0 best quality.

imatrix

Notes: Q4_K_S/Q4_K_M recommended. IQ1_S/IQ1_M for extreme low VRAM. Q6_K for near-original quality.

EXL2

2.5 bpw 3.0 bpw 3.5 bpw 4.0 bpw 4.5 bpw 5.0 bpw 6.0 bpw 8.0 bpw

EXL3

2.0 bpw 2.5 bpw 3.0 bpw 3.5 bpw 4.0 bpw 4.5 bpw 5.0 bpw 6.0 bpw 8.0 bpw

AWQ

Quants

⚠️ Ethical Considerations

This model enhances Broken-Tutu's unalignment:

🚫 Zero moral or ethical constraints on generated content
🔞 Will enthusiastically engage with any NSFW scenario, no matter how extreme
💀 Characters maintain integrity - wholesome characters refuse appropriately, yanderes stab without hesitation
⚖️ Perfectly balanced between character authenticity and user freedom

📜 Performance Notes

🔥 Maintains Omega's intensity with improved narrative coherence
📖 Excels at long-form multi-character scenarios
🧠 Superior instruction following with complex prompts
⚡ Reduced repetition and hallucination compared to v1.1
🎭 Uncanny ability to adapt to subtle prompt nuances
🩸 Incorporates Omega Darker's visceral descriptive power when appropriate
🖼️ Enhanced image understanding capabilities for multimodal interactions

🧑‍🔬 Model Authors

sleepdeprived3 (Training Data & Fine-Tuning)
ReadyArt / Artus / gecfdo (EXL2/EXL3 Quantization)
mradermacher (GGUF Quantization)

☕ Support the Creators

Ko-fi Beaver AI Discord

🔖 License

By using this model, you agree:

To accept full responsibility for all generated content
That you're at least 18+ years old
That the architects bear no responsibility for your corruption