EN-DE parent ➜ EN-NL fine-tuned on creative corpus
Authors: Niek Holter
Thesis: “Transferring Creativity”
Summary
This model starts from Helsinki-NLP’s MarianMT opus-mt-en-fr and is fine-tuned on a 10k-sentence non-creative English–Dutch corpus (Journalistic texts).
It is one of four systems trained for my bachelor’s thesis to study how transfer-learning settings affect MT creativity.
| Parent model | Fine-tune data | BLEU | COMET | Transformer Creativity Score |
|---|
| en-de | Creative | 9.950 | 0.574 | 0.34 |
Intended use
- Research on creative MT and low-resource transfer learning
Training details
- Hardware : NVIDIA GTX 1070 (CUDA 12.1)
- Epochs : Early-stopped ≤ 200 (patience 5)
- LR / batch : 2 e-5 / 16
- Script :
finetuning.py
- Env :
environment.yml
Data
- Non-Creative corpus 10k sentences from DPC Journalistic texts.
- Sentence-level 1:1 alignments; deduplicated to avoid leakage.
See https://github.com/muniekstache/Transfer-Creativity.git for full pipeline.
