🎨 CardProjector-v4
WoonaAi presents...

🚀 Overview
CardProjector is a specialized series of language models, fine-tuned to generate character cards for SillyTavern and for creating characters in general. These models are designed to assist creators and roleplayers by automating the process of crafting detailed and well-structured character cards, ensuring compatibility with SillyTavern's format.
🎨 CardProjector v4 tricks
- Absolute focus on personality development! This version places an absolute emphasis on designing character personalities, focusing on depth and realism. Eight (!) large datasets were collected, oriented towards all aspects of in-depth personality development. Extensive training was also conducted on a dataset of MBTI profiles with Enneagrams from psychology. The model was carefully trained to select the correct personality type according to both the MBTI and Enneagram systems. I highly recommend using these systems (see Usage recommendations); they provide an incredible boost to character realism. I conducted numerous tests with many RP models ranging from 24-70B parameters, and the MBTI profile system significantly impacts the understanding of the character's personality (especially on 70B models), making the role-playing performance much more realistic. You can see an example of a character's MBTI profile here. Currently, version V4 yields the deepest and most realistic characters.
- Reduced likelihood of positive bias! I collected a large toxic dataset focused on creating and editing aggressive, extremely cruel, and hypersexualized characters, as well as transforming already "good harmless" characters into extremely cruel anti-versions of the original. Thanks to this, it was possible to significantly reduce the overall positive bias (especially in Gemma 3, where it is quite pronounced in its vanilla state), and make the model more balanced and realistic in terms of creating negative characters. It will no longer strive at all costs to create a cute, kind, ideal character, unless specifically asked to do so. All you need to do is just ask the model to "not make a positive character, but create a realistic one," and with that one phrase, the entire positive bias goes away.
- Moving to Gemma 3! After a series of experiments, it turned out that this model is ideally suited for the task of character design, as it possesses much more developed creative writing skills and higher general knowledge compared to Mistral 2501 in its vanilla state. Gemma 3 also seemed much more logical than its French competitor.
- Vision ability! Due to the reason mentioned in the point above, you can freely use vision in this version. If you are using GGUF, you can download the mmproj model for the 27B version from bartowski (a vanilla mmproj will suffice, as I didn't perform vision tuning).
- The overall quality of character generation has been significantly increased by expanding the dataset approximately 5 times compared to version V3.
- This model is EXTREMELY sensitive to the user's prompt. So you should give instructions with caution, carefully considering.
- In version V4, I concentrated only on one model size, 27B. Unfortunately, training multiple models at once is extremely expensive and consumes too much effort and time, so I decided it would be better to direct all my resources into just one model to avoid scattering focus. I hope you understand 🙏
💡 Usage Recommendations
Chat Template: Gemma
Balanced output:
Temperature: 0.8
Top-P: 0.92
Rp.Pen: 1.07
Top-K: 100
Rep Pen Range: 360
Rep Pen Slope: 0.7
The character creation process recomendations (updated): In version V4, I want to recommend a new usage method. I strongly recommend generating characters directly in YAML format, rather than starting with natural text first and then converting it to YAML. Gemma 3 turned out to be much better for this use case; you immediately get ready-made, structured characters described in significant detail (in V3, I advised starting the design specifically with natural text, because Mistral and especially Qwen 2.5 had the problem of overly brief character descriptions when generating them directly in YAML).
VERY IMPORTANT THING: As I mentioned earlier, this model is EXTREMELY sensitive to the prompt. When I say EXTREMELY, I really mean EXTREMELY. On one hand, this is a positive aspect—the model does exactly what you tell it to. On the other hand, you need to be very careful about the instructions you give it. You don’t necessarily have to write an overly detailed prompt—just include everything you need to get from the card. Be concise but thorough. The model won’t assume anything extra on behalf of the user—only what the user explicitly asks for. For example, if you want a card with a deeply developed personality, you should specify that in the prompt, or if you want to focus on appearance attributes, you need to emphasize that. That’s how CardProjector V4 works.
MBTI personality profilies: Absolutely must have! This improves the character's personalities several times over the usual description. I strongly recommend using MBTI personality profiles as an add-on to the character card (see the tricks section regarding the MBTI system): first, generate the character card, for example, in YAML format, and when your character is ready, simply ask the model to supplement this work with an MBTI profile and Enneagram, placing it in a separate section within the card. Generally, you can ask the model to create the personality as an MBTI profile right from the first iteration, when you're just starting to generate the character. It's just that the method with the add-on seemed more convenient to me—there's no fundamental difference here.
💬 Examples
MBTI usage (Let's imagine that we've already worked with a character in a chat before)
⚠️ Safety
This model learned on cards for Silly Tavern. I think comments are unnecessary here...
🔧 Technical Specs
Base Model: google/gemma-3-27b-it License: gemma Language: English
- Downloads last month
- 0