DISCOX

Project Overview

DISCOX is a benchmark designed to evaluate the performance of LLMs on discourse-level and expert-level translation tasks.
Unlike existing benchmarks that primarily focus on isolated sentences or general-domain texts, DISCOX emphasizes the ability of models to maintain discourse coherence, terminological precision, and domain-specific accuracy across long-form professional content.

This benchmark covers multiple domains (e.g., social sciences, natural sciences, humanities, applied disciplines, news, domain-specific scenarios, and literature & arts) with a wide range of translation challenges. It enables fine-grained evaluation of LLM outputs using criteria such as accuracy, fluency, and appropriateness.

Getting Started

1. Install Dependencies

Make sure you are using Python 3.9+. Then install the required packages:

pip install -r requirements.txt

2. Configure Environment Variables

Set up your API key and endpoint in the .env file:

JUDGE_API_KEY=your_judgemodel_api_key_here
JUDGE_API_BASE=your_judgemodel_api_base_here
CANDIDATE_API_KEY=your_candidatemodel_api_key_here
CANDIDATE_API_BASE=your_candidatemodel_api_base_here

3. Run Evaluation Tasks

You can run tasks by specifying the target model and the judge model. For example:

python3 run_tasks.py --model openai/gpt4o-2024-11-20 --judgemodel azure/gemini-2.5-pro

Example Use Case

• Model Under Evaluation: openai/gpt4o-2024-11-20
• Judge Model: azure/gemini-2.5-pro
  This configuration runs translation tasks using GPT-4o and evaluates them with Gemini-2.5-Pro under the Metric-S evaluation framework.

---

License

This project is released under the Apache 2.0 License. See LICENSE for details.