LLM Translation Testing

A growing test harness for numerically evaluating machine translation performance on a range of large language models, with a focus on English and Māori translation.

Overview

This project provides a framework to:

• Run translation tasks using various Large Language Models (LLMs), including different quantized versions.
• Evaluate the translation quality using BLEU and chrF metrics.
• Visualize the performance of different models.
• Support for translation between any supported language in the FLORES-200 dataset.
• Facilitate manual comparison of translation outputs.

The primary dataset used is FLORES-200, specifically the devtest sets.

Features

• Multiple Model Support: Test various LLMs, including Hugging Face (HF) and GGUF quantized models (e.g., LLaMA 2 13B Chat and its variants).
• Automated Evaluation: Calculates BLEU and chrF scores for translation outputs.
• Data Visualization: Generates graphs to compare model performance, saved in the output_graphs/ directory.
• Result Persistence: Saves raw translation results in JSON format in the results/ directory.
• Configuration-driven: Experiments are configured through a Python dictionary in the main notebook.
• Colab Integration: Supports running in Google Colab, including GPU acceleration.
• Exploratory Data Analysis: Includes an eda.ipynb notebook for dataset analysis.
• Manual Evaluation Support: Scripts and reports for manual review of translations (see manual_evaluation/ and mini-data-analysis/generate_comparison_report.py).

Directory Structure

.
├── models/                # Potentially for storing or defining models
├── results/               # Stores raw JSON results from translation tasks
├── output_graphs/         # Stores generated graphs and visualizations
├── manual_evaluation/     # Contains reports and data for manual translation comparison
│   └── `manual_evaluation/comparison_report_mri_eng.md`
├── mini-data-analysis/    # Small, ad-hoc analysis scripts
│   └── `mini-data-analysis/generate_comparison_report.py`
│   └── `README.md`
├── `main.ipynb`             # Main Jupyter notebook for running experiments and evaluation
├── `main.py`                # Core Python script with main logic, imported by `main.ipynb`
├── `evaluation.py`          # Script for evaluating translation quality
├── `translation_task.py`    # Defines and manages translation tasks
├── `visualization.py`       # Script for generating result visualizations
├── `prompt.py`              # Handles prompt engineering for LLMs
├── `eda.ipynb`              # Notebook for Exploratory Data Analysis of the dataset
├── `test_model.ipynb`       # Notebook for testing individual models
├── `test_model.py`          # Script for testing individual models
├── `requirements.txt`       # Python dependencies
├── `README.md`              # This file
└── ...                    # Other configuration and project files

Setup

Prerequisites

• Python 3 (tested with Python 3.11, see main.ipynb)
• Pip (Python package installer)
• Git (for cloning, if applicable)
• NVIDIA GPU and CUDA (recommended for GGUF models with ctransformers[cuda])

Dependencies

Install the required Python packages using requirements.txt:

pip install -r requirements.txt

Alternatively, if running in an environment like Google Colab, the main.ipynb notebook includes cells to install specific packages:

# From main.ipynb
# !pip install iso639-lang
# !pip install replicate
# !pip install ctransformers[cuda]>=0.2.24

Dataset

The project uses the Flores200 dataset. Ensure the flores200_dataset directory is populated with the necessary language files (e.g., eng.dev, mri.dev, eng.devtest, mri.devtest). The flores200_dataset.zip file in the repository might contain this dataset.

Google Colab Setup

If using Google Colab, the initial cells in main.ipynb handle mounting Google Drive and changing the working directory:

# filepath: main.ipynb
# ...existing code...
try:
    # Comment out if not using colab
    from google.colab import drive
    drive.mount('/content/drive')

    # Specific for luca's computer
    %cd "/content/drive/Othercomputers/lucas-yoga/Work/Te Taka & Albert LLM project/llm-translation-testing"
    using_colab = True
except:
    print("Not using Google Colab")
    using_colab = False
# ...existing code...

Adjust the %cd path as per your Google Drive structure.

Usage

The main entry point for running experiments is the main.ipynb notebook.

1. Open and Configure: Open main.ipynb in a Jupyter environment or Google Colab.
2. Set Configuration: Modify the config dictionary in the notebook to define which models to run, languages to translate between, and other parameters.

   # filepath: main.ipynb
   # ...existing code...
   config = {
               'run_tasks': False, # Set to True to run new translation tasks
               'run_evaluation': True, # Set to True to evaluate existing/new results
               'viz_output_dir': 'output_graphs',
               'run_bootstrap': False, # Not implemented yet
               'models':
                   [
                   # Example models (refer to main.ipynb for the full list)
                   LLaMA2_13B_Chat_HF,
                   TheBlokeLlama2_13B_chat_Q2_K_GGUF,
                   # ... other models
                   ],
   
               'langs':
                   ['eng_Latn', 'mri_Latn'], # English, Maori
   
               'stop_sequence': '
   

' } # ...existing code... ``` - run_tasks: Set to True to execute translations. If False, it might load existing results if available. - run_evaluation: Set to True to perform BLEU score evaluation. - models: A list of model identifiers/objects to be used for translation. These are typically defined in main.py or imported. - langs: A list of language codes (e.g., eng_Latn for English, mri_Latn for Māori).

3. Run Main Function: Execute the cell calling the main function with the configuration:

   # filepath: main.ipynb
   # ...existing code...
   main(config)
   # ...existing code...
   

4. View Results:

   • Translation outputs are saved as JSON files in the results directory (e.g., results/eng_mri_LLaMA 2 13B Chat - HF.result.json).
   • Evaluation scores (like average BLEU scores) are typically printed in the notebook output.
   • Visualization graphs are saved in the output_graphs directory. An individual scores report will also be generated (e.g., output_graphs/YYYY-MM-DD_HH-MM-SS/scores_report.json).

Evaluation

• Automated: The project uses BLEU and chrF scores for quantitative evaluation, as implemented in evaluation.py. Results are shown in the main.ipynb output.
• Manual: For qualitative analysis, translations can be manually compared. The script mini-data-analysis/generate_comparison_report.py can be used to generate markdown reports like manual_evaluation/comparison_report_mri_eng.md for specific sentences across different models.

Key Files

• main.ipynb: Orchestrates the translation and evaluation tasks.
• main.py: Contains the core logic, model definitions, and the main function.
• evaluation.py: Implements the BLEU score calculation and other evaluation metrics.
• translation_task.py: Manages the process of translating texts with the specified models.
• visualization.py: Generates plots and visual representations of the evaluation results.
• prompt.py: Likely contains functions or classes for formatting prompts fed to the LLMs.
• eda.ipynb: Notebook for performing exploratory data analysis on the Flores200 dataset.