Georgetown University Team for DISRPT 2025 Shared Task

🚀 Try our demo on Hugging Face Spaces:
👉 DeDisco Demo

---

Setup

We use conda for maintaining a python Dev environment and requirements.txt for cataloguing the dependencies

1. Create (or activate) the environment.

conda create -n disrpt python==3.10
conda activate disrpt

2. Install dependencies Dependencies can be installed with

python -m pip install -r requirements.txt

Preparing Task Data

1. Original data processing: Clone the shared task repo (or add as a submodule), navigate to sharedtask2025/util, and run python process_underscore.py. Then copy the output data folder into this repo as data/.
2. Augmented data preparation: Copy the existing augmented_data/ directory into data/ to consolidate all datasets.

Run

This project supports multi-GPU training and single-GPU evaluation via the --mode flag.

Training

To train the model using multiple GPUs (e.g., 4 GPUs):

torchrun --nproc_per_node=4 decoder_w_aug.py --mode train --checkpoint_path output/

• --mode train: Specifies training mode.
• --nproc_per_node=4: Number of GPUs to use.
• --checkpoint_path: Directory to save model checkpoints.

Single-GPU Training

To train the model on a single GPU, you can run the script directly with python:

python decoder_w_aug.py --mode train --checkpoint_path output/

Training Configuration

Our experiments were conducted with an effective batch size of 64. This was achieved using 4 GPUs, a per_device_batch_size of 1, and gradient_accumulation_steps set to 16 (4 GPUs * 1 batch/GPU * 16 steps = 64).

For single-GPU training, set gradient_accumulation_steps to 64 in order to match the effective batch size used in our experiments.

---

Evaluation

To evaluate the model (usually on a single GPU):

python decoder_w_aug.py --mode eval --checkpoint_path output/checkpoint-3827 --res_path res/

• --mode eval: Specifies evaluation mode.
• --checkpoint_path: Directory where the checkpoint is stored, from which the model will be loaded.
  • As training is configured to save only the final model, there will be a single checkpoint folder. This path should point directly to that folder (e.g., output/checkpoint-3827).
• --res_path: Directory to save the prediction results.

To evaluate directly our final model:

python decoder_w_aug.py --mode eval --checkpoint_path JuNymphea/Georgetown-qwen3-4B-finetuned-for-disrpt2025 --res_path res/

Model Checkpoints

• Base model (before fine-tuning)

• Fine-tuned model (trained on task data)