M3.2-36B-Animus-V8.0

Send me your support to help me feed the data beast! also taking comissions for universe specific models

Support on Ko-fi

Quantized Models

The quantized model files are available for download. Click the buttons below to view the files.

Download GGUF Quants → Coming soon → Download EXL3 Quants →

Character Card & Lore Book

For the best roleplaying experience, it is highly recommended to use the provided character card and lore book. These files help guide the model's persona and provide rich, in-universe context.

Download Files →

SillyTavern Sampler Presets

For a seamless setup in SillyTavern, you can download pre-configured sampler presets. These are tuned to provide an optimal balance between creativity and narrative coherence for this model.

Simply download the .json file below and import it into SillyTavern's sampler presets menu.

Download SillyTavern Presets →

Model Description

This is Version 8.0 of the Animus series, a fine-tune of CrucibleLab-TG/M3.2-36b (an upscale of Mistral Small 3.2 24B). This version introduces an experimental approach to structured output while continuing to refine the core roleplaying and DM capabilities of the model.

The goal of this model is to provide the most lore-accurate and immersive conversational experience to date. It can adopt canon character personas with high fidelity, explore alternate timelines from the books, and guide the narrative with new interactive elements.

A surprising outcome of this highly specialized training is that users have reported it is also very capable of general, non-WOF roleplay, making it a more versatile creative partner than previous versions.

Training Details

Training Hardware

This model was trained on 2x H100 SXM GPU.

Training Procedure

A QLoRA (Quantized Low-Rank Adaptation) approach was used for efficient fine-tuning, with an optimized process configured using Axolotl.

Training Data

V8.0 was fine-tuned on a high-quality dataset of 3,200 examples with several key improvements:

• Experimental Structured Output: V8.0 was trained with a custom tokenizer and vocabulary in an attempt to teach the model to wrap its multiple-choice suggestions in tags. Note: This feature is highly experimental and rarely works as intended. However, the underlying training has resulted in a very coherent and high-quality model for general roleplay.
• Canon-Centric Scenarios: All roleplay scenarios are based on pivotal events from the Wings of Fire book series, exploring "what-if" outcomes. (e.g., What if Darkstalker didn't kill Arctic at that moment?). This ensures deep and lore-consistent interactions.
• Canon-Only Characters: The model was trained exclusively on canon characters from the books. AI-generated characters have been removed from the training data (except for the user's persona), leading to more authentic character portrayals.
• Dungeon Master (DM) Enhancement: The model's ability to act as a Dungeon Master has been further enhanced, prompting the user with multiple-choice actions to drive the story forward. For example: You arrive in front of Queen Scarlet. What do you do? A)... B)... C)...
• Improved Data Cleaning: The dataset underwent a rigorous cleaning process to remove formatting artifacts from previous versions, such as **scene transitions**, resulting in a cleaner and more natural narrative style.
• Refined Turn Structure: Addressed an issue where consecutive AI turns appeared in the dataset, leading to a healthier learning curve and more natural conversational flow.

Intended Use & Limitations

• Intended Use: The primary purpose of this model is for creative and roleplaying within the Wings of Fire universe. However, user feedback indicates it is also highly effective for general-purpose roleplaying.
• Limitations & Quirks:
  • Experimental Features: The custom tag functionality rarely works. The model may occasionally attempt to use it, but users should not expect reliable structured output in this format. Despite this, the model's overall quality remains very high.
  • Performance on tasks outside of its training domain (general knowledge, coding, etc.) is not guaranteed and will likely be poor.
  • Versatility: While specifically tuned for Wings of Fire, the model has proven to be very capable of performing normal roleplay with other settings and characters.
  • The model may "hallucinate" or generate plausible but non-canonical information, especially when pushed outside the established "what-if" scenarios.
  • Content: The training data includes mature and darker themes from the Wings of Fire series, such as conflict, character death, and moral ambiguity. The model is capable of generating content reflecting these themes. As always, it is up to the user what they do with it.
  • Formatting: Training data was cleaned to remove narrative artifacts like **scene transitions**. The model should now produce cleaner prose.
  • Safety: This model has not undergone additional safety alignment beyond what was included in its `M3.2-36b` base model. Standard responsible AI practices should be followed.

Acknowledgements

• Credit to Mistral AI and CrucibleLab-TG for the powerful M3.2-36b base model.
• Credit to Google for the Gemini Pro model, used in dataset generation.
• Credit to Evan Armstrong for Augmentoolkit, an invaluable tool for dataset creation.