Please read this

Please read this

Summary:

  • New Sugoi translation models are coming out, I'll release them in Sugoi Toolkit V13 to Patreon supporters and open-source them to the public in V14
  • I'll release the local version of Premium Manga Translator in V13.
  • You can expect V13 around June
  • I'll raise the base tier price of Patreon from 3 dollars to 5 dollars to address growing cost and also integrate more LLM features inside Sugoi Toolkit.

The return of Sugoi models:

Above is the translation benchmark for an entire visual novel from start to end (Harmonia from Visual Arts/Key with 4569 dialogues lines)

I'll public all model's translation result soon.

Sugoi 32B and 14B are a massive leaps over Sugoi Offline V4. These two models completely solve MTL awkward wording issues, making translation sounds just like humans.

The final barriers are now pronouns/context accuracy and styling.

Here is a sample translation from Sugoi 32B

For those curious, "similarity" score is used for measuring accuracy in intentions. "BLEU" and "chrF" are for measuring word similarity (style).

These two styles of scores are important because let's say ChatGPT is great at understanding a VN sentence so it can have a high similarity score, but if it is not weeb enough, it's translation may sounds like a quote from Shakespeare which will have low BLEU and chrF score.

Let's solve manga translation together:

I'll release locally Premium Manga Translator. You'll be able to access all of it's source code as well as run it offline directly inside Sugoi Toolkit.  

In the future, I'll further develop it to be an AI powered editor so that you, the readers, can take actions to edit and share your efforts easily.

To level up, Sugoi Toolkit needs additional support:

Sugoi Toolkit is growing in complexity. As of V12, there are 21 options on the menu screen compared to just two on V1.

Everyday, there are at least 10 new users joining, and some new users and old users will always mysterious find new bugs that I never know about.

The training of the 14B and 32B models costed a total of $10,000. Half from me, half from Cirno, the developer of Mtool. (btw, Cirno is probably the nicest guy in the industry, he told me he just wanted to see how far we can push Sugoi models)

On June, I'll increase the base Patreon tier from $3 to $5. (I'll do it soon which will take 31 days to be effective)

The additional money will be use to level up the research and development of Sugoi Toolkit.

I'll also integrate LLM features as first class citizens and give all Patreon supporters Sugoi coins to instantly access them to read/translate the content you like.

For example, I recently benchmarked various LLMs for OCR. ChatGPT are getting close to 100% accuracy on Japanese text recognition. (probably the same for Korean and Chinese). The cost of these models will only get lower with time while their performance will get better. The future is almost here and we'll have to pay a bit for it.