This is the model developped by the Melodi team for the open track of the Disrpt2025 shared track, for the task of discourse relation labelling. It's based on Qwen3-4B quantized, fine-tuned with the Lora adapter on relation labeling, using the unsloth library. The training script is provided (unsloth_classif_SFT), along with a notebook (Test_inference) to evaluate the model on disrpt2025 datasets (beware that some of the datasets are not directly provided by disrpt as they are copyrighted).

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.