🦖 T-Ragx

Enhancing Translation with RAG-Powered Large Language Models

T-Ragx Demo:

TL;DR

Overview

• Open-source system-level translation framework
• Provides fluent and natural translations utilizing LLMs
• Ensures privacy and security with local translation processes
• Capable of zero-shot in-task translations

Methods

• Utilizes QLoRA fine-tuned models for enhanced accuracy
• Employs both general and in-task specific translation memories and glossaries
• Incorporates preceding text in document-level translations for improved context understanding

Results

• Combining QLoRA with in-task translation memory and glossary resulted in ~45% increase in aggregated WMT23 translation scores, benchmarked against the Mistral 7b Instruct model
• Demonstrated high recall for valid translation memories and glossaries, including previous translations and character names
• Surpassed the performance of the native TowerInstruct model in three (Ja<->En, Zh->En) out of the four WMT23 language direction tested
• Outperformed DeepL in translating the Japanese web novel "That Time I Got Reincarnated as a Slime" into Chinese using in-task RAG
  • Japanese to Chinese translation improvements:
    • +29% sacrebleu
    • +0.4% comet22

👉See the write-up for more details📜

Getting Started

Install

Simply run:

pip install t-ragx

or if you are feeling lucky:

pip install git+https://github.com/rayliuca/T-Ragx.git

Elasticsearch

See the wiki page instructions

Note: you can access preview read-only T-Ragx Elasticsearch services at https://t-ragx-fossil.rayliu.ca and https://t-ragx-fossil2.rayliu.ca (But you will need a personal Elasticsearch service to add your in-task memories)

Environment

(Recommended) Conda / Mamba

Download the conda environment.yml file and run:

conda env create -f environment.yml

## or with mamba
# mamba env create -f environment.yml

Which will crate a t_ragx environment that's compatible with this project

pip

Download the requirment.txt file and run:

Use your favourite virtual environment, and run:

pip install -r requirment.txt

Examples

Initiate the input processor:

import t_ragx

# Initiate the input processor which will retrieve the memory and glossary results for us
input_processor = t_ragx.Processors.ElasticInputProcessor()

# Load/ point to the demo resources
input_processor.load_general_glossary("https://t-ragx-public.s3.us-west-004.backblazeb2.com/t-ragx-public/glossary")
input_processor.load_general_translation(elasticsearch_host=["https://t-ragx-fossil.rayliu.ca", "https://t-ragx-fossil2.rayliu.ca"])

Using the llama-cpp-python backend:

import t_ragx

# T-Ragx currently support 
# Huggingface transformers: MistralModel, InternLM2Model
# Ollama API: OllamaModel
# OpenAI API: OpenAIModel
# Llama-cpp-python backend: LlamaCppPythonModel
mistral_model = t_ragx.models.LlamaCppPythonModel(
    repo_id="rayliuca/TRagx-GGUF-Mistral-7B-Instruct-v0.2",
    filename="*Q4_K_M*",
    # see https://huggingface.co/rayliuca/TRagx-GGUF-Mistral-7B-Instruct-v0.2
    # for other files
    chat_format="mistral-instruct",
    model_config={'n_ctx':2048}, # increase the context window
)

t_ragx_translator = t_ragx.TRagx([mistral_model], input_processor=input_processor)

Translate!

t_ragx_translator.batch_translate(
    source_text_list,  # the input text list to translate
    pre_text_list=pre_text_list,  # optional, including the preceding context to translate the document level
    # Can generate via:
    # pre_text_list = t_ragx.utils.helper.get_preceding_text(source_text_list, max_sent=3)
    source_lang_code='ja',
    target_lang_code='en',
    memory_search_args={'top_k': 3}  # optional, pass additional arguments to input_processor.search_memory
)

Models

Note: you could use any LLMs by using the API models (i.e. OllamaModel or OpenAIModel) or extending the t_ragx.models.BaseModel class

The following models were finetuned using the T-Ragx prompts, so they might work a bit better than some of the off-the-shelve models with T-Ragx

QLoRA Models:

Data Sources

All of the datasets used in the project