Mistral 24B
Mistral 24B is one of CrushOn's LLMs available for use by user. One of the older but reliable models available in CrushOn, the Mistral 24B officially focuses on faster response speed, with a tendency to provide concise replies. However, despite the official description, after the many iterations of improvements that the model has had, it actually performs quite well in delivering moderate to longer lengths of responses.
| Free 8K | Pro 8K | Pro 16K | Pro 24K |
|---|---|---|---|
| ✓ | ✓ | ✓ | ✕ |
From here on out, the content may not be fully accurate to each user and act only as general reference. Each user often could see highly different experiences as there are many variables that affect the behavior and quality of the models. From profile cards, the style in which the user converse with the character, the set parameters for the model, and of course the personality of the characters. And periodically, even minimal adjustments made to the model by the developers could also substantially change its behavior whether intended or not.
Strengths and Flaws
Mistral is a reliable and fast storyteller, being able to accommodate for SFW and NSFW scenarios with equal quality. It will be able to pick up on cues and progress plot accordingly. It is expressive, creative, and accurate to character personalities. Being a cheaper model, it can also serve as a great alternative to the more popular models, especially when used in English. The model is also obedient, understanding and following through on profile cards and custom instructions with no delay and very low chance of inaccuracy.
However, with the speed and stability it provides, it comes with some deterring caveats, especially for first time users.
- Mainly its greatest flaw would be the initial 'setup' required to keep it going and consistent throughout the rest of the conversation. This 'setup' is mostly exhibited through inconsistent formatting in initial stages of the chat where the model may respond by misplacing or forgetting asterisks in certain places.
- Perhaps what is less considered an issue and more to a style flaw, is that Mistral 24B would often incite the actions of {{user}} or continue it without additional input. Often case, it will be a natural progression, but it is understandable that it may not suit the taste of many users to have their actions narrated for them.
Common Issues with the Model
Besides the most jarring of all the issues with the tendency to have initial broken formatting, the model Mistral 24B most often comes with notable issues.
- The model misplaces, forgets, or adds additional asterisks: As mentioned above, Mistral's greatest weakness and issue lies in the fact that, most of the time, the model would start conversations with broken formatting. However, with patience, this isn't quite as big of an issue as you would think. By simply manually editing the formatting into one that the user prefers, the model will eventually naturally learn the style and be more uniform. Though for how many messages one should manually edit it, varies depending on the character and style preference.
- In trying to extend response lengths, the model instead becomes repetitive: From time to time, the model will try to extend or make most use of its token limit, but in doing so it becomes lazy and simply 'copy and paste' a passage or phrase from the same response. Regenerating sometimes help, but a more consistent method to alleviate this symptom is to tick the "Reduce Repetition" option in Instructions.
- The model may cut off for free users from time to time: Lastly, while the model is officially described as concise, for free users with 225 token limit, the model might actually cut off from overly long responses instead. Adding in a custom instruction such as "{{char}} uses a maximum of 350 characters" would often fix this issue.
Example of Use
On the right, you could see the example of the model being used regularly with the OC Yumi by Since.