AlexBefest/CardThinker-32B-v3

🌙 CardThinker-v3

WoonaAi presents...

🚀 Overview

CardThinker is an offshoot of the CardProjector series, but with the possibility of reasoning. CardProjector and CardThinker 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.

✨ CardThinker v3 tricks

• Due to its ability to think - CardThinker V3 32B works through characters much deeper, more thoroughly, and logically.
• Because of the peculiarities of the base model QwQ, this model has VERY well-developed creative writing skills (to the extent that I found the most usable temperature range to be 0.1-0.3).
• Among the major drawbacks - high positive bias; CardThinker V3 tries to create an ideal friendly character. Even if you write "be sure to come up with negative personality traits" (or something along those lines), it will come up with traits that are too soft and harmless. To create a negative character, you have to fully come up with such traits yourself and then make the model develop them. Because of all this, in my opinion, CardProjector 24B V3 produces more vivid negative characters.
Use CardThinker if:
• You need a very logical and deeply developed character card.
• You need crazy creative ideas.
• You need a very good character editor (CardThinker significantly outperforms CardProjector in editing cards and following instructions).
Use CardProjector if:
• You need balanced, more realistic characters without over-the-top expressiveness, but much simpler and less detailed.
• You need high-quality "First Messages" (CardProjector excels at creating dialogue examples for characters).

💡 Usage Recommendations

Reasoning mode:

This model can be used in two modes: reasoning and standard. The standard mode is enabled by default. However, if you need the advanced reasoning mode with its chain-of-thought capability (I strongly recommend using it, as it significantly improves the quality of the generated cards), you must initiate the model's response starting with the following line:

<think>
Okay, user is

In SillyTavern, you can find this setting under "AI Response Formatting - Start Reply With" and simply paste the line above.

In other GUIs that lack a similar feature, you can try this workaround:

1. Send your request.
2. Receive the first few tokens from the model.
3. Stop the generation.
4. Delete what the model has generated.
5. Manually edit its response by inserting the line <think>\nOkay, user is at the very beginning.
6. Prompt the model to continue generating its response from there.

This alternative method has also worked for me, although it is more cumbersome.

Chat history: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content! (From official Qwen QwQ usage tips)

Chat Template: ChatML

Balanced output:
Temperature: 0.1-0.3
Top-P: 0.92
Rp.Pen: 1.07
Top-K: 100
Rep Pen Range: 360
Rep Pen Slope: 0.7

Ready-to-use SillyTavern preset: here

The character creation process: Based on my tests, I would recommend the following approach. To create a well-developed and structured character, I suggest first asking the model to generate the character in a standard, natural format (meaning you shouldn't request formats like YAML or JSON right away), allowing it to describe the character in plain, understandable text. Then, if needed, ask for any necessary adjustments. Once you're satisfied with the result, request the final version to be converted into YAML format. Why YAML? It's an ideal format for structuring and summarizing a character from your chat story. This format is human-readable, and its clear structure is very well processed by RP models (from my tests, it’s even better in some ways than XML). You can simply copy the entire YAML output and paste it into the Description field in Silly Tavern. Alternatively, you can ask the model to convert the resulting card into JSON while leaving the YAML description untouched. I have found this method of using CardProjector v3 to be the most effective.

⚠️ Safety

This model learned on cards for Silly Tavern. I think comments are unnecessary here...

🔧 Technical Specs

Base Model: Qwen/QwQ-32B License: Apache-2.0 Language: English

© 2024 WoonaAI | 🤗 Hosted on Hugging Face