Comparison of Prompt Variations on Poetry Quality Using a Fine-Tuned GPT-2 Model

DOI

10.2991/978-94-6239-828-3_12How to use a DOI?

Keywords

GPT-2; Fine-Tuning; Poetry Generation; Prompt Variation; Indonesian Language

Abstract

Natural Language Generation (NLG) enables generative models such as GPT-2 for Indonesian poetry generation, yet challenges re- main in fluency and diversity for low-resource languages. This study analyzes the effects of prompt variations (zero-shot, one-shot, and few-shot), hyperparameters, and fine-tuning on generated poetry quality. Three models were evaluated: GPT-2 Small Indonesia (pre-trained), GPT-2 Small Indonesia Fine-Tuning Poem (pre-trained), and GPT-2 FT (further fine-tuned on 125 poems by Goenawan Mohamad). Perplexity and distinct n-gram (distinct-1, distinct-2) were used to measure fluency and lexical diversity. The results reveal a pronounced trade-off: GPT-2 FT achieved the lowest perplexity (1.907) under one-shot prompting, which suggests highest predictive confidence, but also produced the lowest lexical diversity (distinct-1 = 0.1146, distinct-2 = 0.1809), a classic example of degenerative low-entropy output. Conversely, the highest distinct unigram (0.9169) and bigram (0.9809) scores were obtained by pre-trained models. These findings demonstrate that prompt variation, hyperparameter tuning, and fine-tuning significantly influence the fluency-diversity trade-off.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - I Wayan Ivan Zenatmaja
AU  - Tirana Noor Fatyanosa
AU  - Prima Zulvarina
AU  - Heru Nurwarsito
PY  - 2026
DA  - 2026/11/04
TI  - Comparison of Prompt Variations on Poetry Quality Using a Fine-Tuned GPT-2 Model
BT  - Proceedings of the 2026 11th International Conference on Sustainable Information Engineering and Technology (SIET 2026)
PB  - Atlantis Press
SP  - 133
EP  - 142
SN  - 2589-4900
UR  - https://doi.org/10.2991/978-94-6239-828-3_12
DO  - 10.2991/978-94-6239-828-3_12
ID  - Zenatmaja2026
ER  -