Hi everyone,

MTW can read GGUF models even if they are not LLAMA type, for example GPT. However, the results will most likely be strange because MTW is configured to follow the LLAMA chat template since version 3. LLAMA 3 models can be used but it is not good for languages ​​other than English. LLAMA 3.1 already includes Spanish, German, French, Italian, Portuguese, Hindi and Thai. That is why it is the best fit.

A model with more bits is more accurate (obviously) but the memory consumption and execution time will be higher. There you have to test which model is well tolerated by your computer.

In theory, at this moment if there is a LLAMA 3.1 model trained for another language, MTW could run in that language. I have tested with Korean and Chinese and it does well, because there are 3.1 models for that (unfortunately I haven't found any Japanese).

Interesting things to try:

+ is there an uncensored Llama 3.1 better than Bartowski? The one currently used as default is uncensored but maybe there is one that does it better.

+ Recently LLAMA 3.2 came out, very optimized, focused on embedded systems and mobile, would it work well on MTW?

Regarding the size of the personality txt, although the context has a range set at 8000 tokens (LLAMA 3.1 supports more than that, but that is used for example to summarize books) it is equivalent to more or less 7500 words. I think that a thousand words is enough.

The script_expressions file is the one that contains the script for certain parts of the game. Right now it only has the OpenDoor, which is when a certain number of interactions is reached, the signal to open the door is sent internally to UE4, and a predefined question is sent to LLAMA.

Regarding the voices, I haven't tried it, but this page seems like a good option if you want a nicer woman voices than the ones provided by Windows: https://www.cereproc.com/en/storesapi

as long as it's 64-bit, MTW should recognize it. If anyone has the chance to try it and tell us if it works that would be great to be more sure.

The next update will focus on:

+ Screen options: resolution, antialiasing, etc.
+ Memory capacity: MTW will save a file of the conversation you had so that in subsequent sessions the context is recovered and it "remembers you". Now it is more complex to implement because it would be a file for each personality of the android, a Manager class must be developed.