ALMA 7B Ja V2

🚀 New Translation Model Released

The C3TR-Adapter is the QLoRA adapter for google/gemma-7b. Despite the 4-bit quantization, the GPU memory requirement has increased to 8.1 GB. However, it can be run on the free version of Colab, and the performance is significantly improved!

✨ Features

webbigdata/ALMA-7B-Ja-V2

The ALMA-7B-Ja-V2 is a machine translation model capable of translating between Japanese and English, as well as between the following languages, although its primary focus is on Japanese-English and English-Japanese translation:

• German (de) and English (en)
• Chinese (zh) and English (en)
• Icelandic (is) and English (en)
• Czech (cs) and English (en)

This model builds on the previous one (ALMA-7B-Ja) by adding further learning, thereby enhancing its performance.

Benchmark Results

The following three metrics were used to evaluate the translation performance. A higher score indicates better performance:

BLEU

A metric that assesses the similarity between the translated text and the original text. However, it mainly focuses on word frequency and may not adequately evaluate word order accuracy or sentence fluency.

chrF++

A method for evaluating translation accuracy based on character combination matches and word order. Its drawback is that it may not be suitable for evaluating long sentences.

comet

A tool that uses machine learning models to automatically evaluate translation quality. It is said to be similar to human subjective evaluation, but since it is based on machine learning, it highly depends on the data used for training by the original model.

Comparison with NLLB-200

The benchmark results compared to Meta's NLLB-200 series of super multilingual machine translation models, which support over 200 languages, are as follows:

Comparison with Previous Model (ALMA-7B-Ja)

ALMA-7B-Ja-V2

Comparison with Real-World Applications

The results of comparing ALMA-7B-Ja-V2 with real-world applications across various text genres are as follows:

Government Official Announcements

Classical Literature

Fanfiction

💻 Usage Examples

Basic Usage

Using Colab, Google's free web tool, you can easily verify the performance of ALMA_7B_Ja_V2. Sample Code For Free Colab

📚 Documentation

Other Versions

llama.cpp

The main purpose of llama.cpp is to run the LLaMA model using 4-bit integer quantization on a MacBook. Although performance is somewhat reduced with 4-bit quantization, webbigdata-ALMA-7B-Ja-V2-gguf, created by mmnga, can be used to run this model on Mac, Windows, and Linux without a GPU. Here is Colab(without GPU) sample code

GPTQ

GPTQ is a technique (called quantization) that reduces model size. ALMA-7B-Ja-V2-GPTQ-Ja-En is a quantized GPTQ version, which reduces model size (3.9 GB) and memory usage and increases speed. However, performance is slightly reduced, and the ability to translate into languages other than Japanese and English should be significantly reduced. Sample Code For Free Colab webbigdata/ALMA-7B-Ja-V2-GPTQ-Ja-En

If you want to translate the entire txt file at once, try Colab below. ALMA_7B_Ja_GPTQ_Ja_En_batch_translation_sample

Model Details

ALMA (Advanced Language Model-based trAnslator) is an LLM-based translation model that 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.

📄 License

The license for this model is llama2.