inference: false
language:
- ja
- en
- de
- is
- zh
- cs
webbigdata/ALMA-7B-Ja-V2
ALMA-7B-Ja-V2は日本語から英語、英語から日本語の翻訳が可能な機械翻訳モデルです。 The ALMA-7B-Ja-V2 is a machine translation model capable of translating from Japanese to English and English to Japanese.
ALMA-7B-Ja-V2は以前のモデル(ALMA-7B-Ja)に更に学習を追加し、性能を向上しています。
The ALMA-7B-Ja-V2 adds further learning to the previous model (ALMA-7B-Ja) and improves performance.
日本語と英語間に加えて、このモデルは以下の言語間の翻訳能力も持っています。
In addition to translation between Japanese and English, the model also has the ability to translate the following four languages.
- ドイツ語 German(de) and 英語 English(en)
- 中国語 Chinese(zh) and 英語 English(en)
- アイスランド語 Icelandic(is) and 英語 English(en)
- チェコ語 Czech(cs) and 英語 English(en)
ベンチマーク結果
Meta社の200言語以上の翻訳に対応した超多言語対応機械翻訳モデルNLLB-200シリーズと比較したベンチマーク結果は以下です。
Benchmark results compared to Meta's NLLB-200 series of super multilingual machine translation models, which support translations in over 200 languages, are shown below.
NLLB-200
| Model Name | file size | E->J chrf++/F2 | E->J comet | J->E chrf++/F2 | J->E comet |
|---|---|---|---|---|---|
| NLLB-200-Distilled | 2.46GB | 23.6/- | - | 50.2/- | - |
| NLLB-200-Distilled | 5.48GB | 25.4/- | - | 54.2/- | - |
| NLLB-200 | 5.48GB | 24.2/- | - | 53.6/- | - |
| NLLB-200 | 17.58GB | 25.2/- | - | 55.1/- | - |
| NLLB-200 | 220.18GB | 27.9/33.2 | 0.8908 | 55.8/59.8 | 0.8792 |
previous our model(ALMA-7B-Ja)
| Model Name | file size | E->J chrf++/F2 | E->J comet | J->E chrf++/F2 | J->E comet |
|---|---|---|---|---|---|
| webbigdata-ALMA-7B-Ja-q4_K_S | 3.6GB | -/24.2 | 0.8210 | -/54.2 | 0.8559 |
| ALMA-7B-Ja-GPTQ-Ja-En | 3.9GB | -/30.8 | 0.8743 | -/60.9 | 0.8743 |
| ALMA-Ja(Ours) | 13.48GB | -/31.8 | 0.8811 | -/61.6 | 0.8773 |
ALMA-7B-Ja-V2
| Model Name | file size | E->J chrf++/F2 | E->J comet | J->E chrf++/F2 | J->E comet |
|---|---|---|---|---|---|
| ALMA-7B-Ja-V2-GPTQ-Ja-En | 3.9GB | -/33.0 | 0.8818 | -/62.0 | 0.8774 |
| ALMA-Ja-V2(Ours) | 13.48GB | -/33.9 | 0.8820 | -/63.1 | 0.8873 |
| ALMA-Ja-V2-Lora(Ours) | 13.48GB | -/33.7 | 0.8843 | -/61.1 | 0.8775 |
様々なジャンルの文章を実際のアプリケーションと比較した結果は以下です。
Here are the results of a comparison of various genres of writing with the actual application.
政府の公式文章 Government Official Announcements
| e->j chrF2++ | e->j BLEU | e->j comet | j->e chrF2++ | j->e BLEU | j->e comet | |
|---|---|---|---|---|---|---|
| ALMA-7B-Ja-V2-GPTQ-Ja-En | 25.3 | 15.00 | 0.8848 | 60.3 | 26.82 | 0.6189 |
| ALMA-Ja-V2 | 27.2 | 15.60 | 0.8868 | 58.5 | 29.27 | 0.6155 |
| ALMA-7B-Ja-V2-Lora | 24.5 | 13.58 | 0.8670 | 50.7 | 21.85 | 0.6196 |
| gpt-3.5 | 34.6 | 28.33 | 0.8895 | 74.5 | 49.20 | 0.6382 |
| gpt-4.0 | 36.5 | 28.07 | 0.9255 | 62.5 | 33.63 | 0.6320 |
| google-translate | 43.5 | 35.37 | 0.9181 | 62.7 | 29.22 | 0.6446 |
| deepl | 43.5 | 35.74 | 0.9301 | 60.1 | 27.40 | 0.6389 |
二次創作 Fanfiction
| e->j chrF2++ | e->j BLEU | e->j comet | j->e chrF2++ | j->e BLEU | j->e comet | |
|---|---|---|---|---|---|---|
| ALMA-7B-Ja-V2-GPTQ-Ja-En | 27.6 | 18.28 | 0.8643 | 52.1 | 24.58 | 0.6106 |
| ALMA-Ja-V2 | 20.4 | 8.45 | 0.7870 | 48.7 | 23.06 | 0.6050 |
| ALMA-7B-Ja-V2-Lora | 23.9 | 18.55 | 0.8634 | 55.6 | 29.91 | 0.6093 |
| gpt-3.5 | 31.2 | 23.37 | 0.9001 | - | - | 0.5948 |
| gpt-4.0 | 30.7 | 24.31 | 0.8848 | 53.9 | 24.89 | 0.6163 |
| google-translate | 32.4 | 25.36 | 0.8968 | 58.5 | 29.88 | 0.6022 |
| deepl | 33.5 | 28.38 | 0.9094 | 60.0 | 31.14 | 0.6124 |
Other Version
ALMA-7B-Ja-V2-GPTQ-Ja-En
GPTQ is quantized(reduce the size of the model) method and ALMA-7B-Ja-V2-GPTQ has GPTQ quantized version that reduces model size(3.9GB) and memory usage.
But the performance is probably lower. And translation ability for languages other than Japanese and English has deteriorated significantly.
Sample Code For Free Colab webbigdata/ALMA-7B-Ja-V2-GPTQ-Ja-En
If you want to translate the entire file at once, try Colab below.
ALMA_7B_Ja_GPTQ_Ja_En_batch_translation_sample
ALMA (Advanced Language Model-based trAnslator) is an LLM-based translation model, which adopts a new translation model paradigm: it begins with fine-tuning on monolingual data and is further optimized using high-quality parallel data. This two-step fine-tuning process ensures strong translation performance. Please find more details in their paper.
@misc{xu2023paradigm,
title={A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models},
author={Haoran Xu and Young Jin Kim and Amr Sharaf and Hany Hassan Awadalla},
year={2023},
eprint={2309.11674},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
Original Model ALMA-7B. (26.95GB)
Prevous Model ALMA-7B-Ja. (13.3 GB)
about this work
- This work was done by : webbigdata.