TLAssist-llama2-7b-experiment-qlora

TLAssist is a qlora trained for 2 epochs using all the lines of the final chapter of the Visual Novel "Rewrite" in groups of "Original Text" (Japanese), "Baseline Translation" (MTL), "Improved Translation" (Fan TL). The purpose of this experiment being to find out if it's possible to create a model that can aid improve the quality of MTL.

Prompt

Original Text: {Japanese Text}
Baseline Translation: {MTL}
Improved Translation:

For multi-turn, each translation group is separated by ---.

v1

The results I got from v1 seem to be promising, however, It still isn't ideal, some examples of things that I didn't like are:

Original Text: 「…う、うん」
Baseline Translation: "…No"
Improved Translation: 「...no.」 // The model seems to trust the baseline too much.

Original Text: その一　いつもと違う髪形に気がつくこと
Baseline Translation: One thing is noticing a different hairstyle than usual
Improved Translation: One: you must notice when I've changed my hairstyle
---
Original Text: その二　ちゃんと靴までみることいいね？
Baseline Translation: The second thing, it's good to pay attention to the shoes as well, right?
Improved Translation: Two: make sure you look at everything from head to toe. // The model seems to omit things for no reason. (in this case the "right?")