---
tags:
  - conversational
---

DialoGPT-small finetuned in the Maptask Corpus. The [repository](https://github.com/KonstSkouras/Maptask-Corpus/tree/develop) with additionally pre-processed Maptask dialogues of concatenated utterances per speaker, 80/10/10 train/val/test split and metadata, is a fork of the [Nathan Duran's repository](https://github.com/NathanDuran/Maptask-Corpus). For finetuning the `train_dialogpt.ipynb` notebook from Nathan Cooper's [Tutorial](https://nathancooper.io/i-am-a-nerd/chatbot/deep-learning/gpt2/2020/05/12/chatbot-part-1.html) was used to finetune the model with slight modifications in Google Collab.

History dialogue context = 5. Number of utterances: 14712 (train set), 1964 (test set), 2017 (val set). Fine-tuning for 3 epochs with batch size 2.

Evaluation perplexity in Maptask from 410.7796 (pre-trainded model) to 19.7469 (fine-tuned model).