JinaJudge Training Code

This repository contains the code used to train JinaJudge, a model designed to replicate GPT-4-1106-Preview judgments in the Russian LLM Arena for more cost-effective model evaluations.

Human judgments are expensive and time-consuming, and while GPT-4 can act as a proxy, it remains costly. JinaJudge reduces costs by replicating these judgment patterns in a lightweight, scalable model.

Installation

Clone this repository and install the package:

git clone https://github.com/oKatanaaa/jina-judge
pip install -e .

Usage

To train the model, follow these general steps:

1. Prepare Data: Ensure you have datasets formatted as .jsonl files (see the example for data structure).
2. Configure Training: Customize the training parameters in the provided config.yaml file.
3. Run Training: Use the provided bash script in the scripts folder to start training:

bash run.sh config.yaml

The script will save the best-performing model weights, training logs, and outputs in the specified directories. For more details, refer to the example config and dataset structure in the repository.

NOTE: running the code requires an Ampere GPU or newer due to FlashAttention requirement in the JINA encoder.