CRAFT – Custom KODIS Implementation

This repository contains our tailored implementation of the Conversational Recurrent Architecture for ForecasTing (CRAFT) neural model, originally introduced in the EMNLP 2019 paper Trouble on the Horizon: Forecasting the Derailment of Online Conversations as they Develop to the dispute resolution domain. We pre-train the CRAFT model architecture with a custom corpus of CaSiNo, Deal no Deal, and KODIS dialogs and finetune on the KODIS dataset to research whether we can learn unsupervised representation of conversational dynamics in negotiation-based dialogues and expoit the structure via supervised learning in fine-tune for predicting for outcomes in Dispute resolution (KODIS).

---

1. Prerequisites

• Python 3.8+
• PyTorch 1.10+
• Ray 2.x (Tune & AIR)
• MLflow 2.x
• ConvoKit 3.x scikit-learn, pandas, matplotlib, NLTK

---

2. Fine-tuning on KODIS with Ray Tune & MLflow

Everything is orchestrated in src/runners/raytune.py. By default it will:

• Perform k-fold cross-validation on your train split of KODIS.

• Report per-fold batch losses and mean validation metrics each epoch via tune.report.

• Use ASHAScheduler to early-stop underperforming hyperparameter trials.

• Log all parameters, metrics, and model artifacts to MLflow.

---