NLLB 1.3B fine-tuned on Japanese to English Light Novel translation

This model was fine-tuned on light and web novel for Japanese to English translation.

It can translate sentences and paragraphs up to 512 tokens.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("thefrigidliquidation/nllb-jaen-1.3B-lightnovels")
model = AutoModelForSeq2SeqLM.from_pretrained("thefrigidliquidation/nllb-jaen-1.3B-lightnovels")

generated_tokens = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.lang_code_to_id[tokenizer.tgt_lang],
    max_new_tokens=1024,
    no_repeat_ngram_size=6,
).cpu()

translated_text = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]

Generating with diverse beam search seems to work best. Add the following to model.generate:

num_beams=8,
num_beam_groups=4,
do_sample=False,

Glossary

You can provide up to 10 custom translations for nouns and character names at runtime. To do so, surround the Japanese term with term tokens. Prefix the word with one of <t0>, <t1>, ..., <t9> and suffix the word with </t>. The term will be translated as the prefix term token which can then be string replaced.

For example, in マイン、ルッツが迎えに来たよ if you wish to have マイン translated as Myne you would replace マイン with <t0>マイン</t>. The model will translate <t0>マイン</t>、ルッツが迎えに来たよ as <t0>, Lutz is here to pick you up. Then simply do a string replacement on the output, replacing <t0> with Myne.

Honorifics

You can force the model to generate or ignore honorifics.

# default, the model decides whether to use honorifics
tokenizer.tgt_lang = "jpn_Jpan"
# no honorifics, the model is discouraged from using honorifics
tokenizer.tgt_lang = "zsm_Latn"
# honorifics, the model is encouraged to use honorifics
tokenizer.tgt_lang = "zul_Latn"