Korbeeck, Ruben (2026) Investigating the Stability Gap in Large Language Models During Incremental Instruction Tuning. Bachelor's Thesis, Artificial Intelligence.
|
Text
BScThesisRubenKorbeeck.pdf Download (5MB) | Preview |
Abstract
This thesis investigates whether stability-gap behaviour occurs during joint-incremental instruction tuning of large language models. In this setting, a model is first fine-tuned on one instruction-following task and then further trained on a joint dataset containing examples from both the original task and a newly introduced task. Unlike standard catastrophic forgetting, where previous-task performance may decline persistently, a stability gap refers to a temporary decrease in previous-task performance after the new task is introduced, followed by partial or complete recovery. Experiments were conducted using Pythia-14M, Pythia-410M, and Pythia-6.9B on SST-2 and QNLI. After introducing the second task, performance on the original task temporarily decreased before recovering later in training, showing that stability-gap behaviour can occur in this setting. A two-way ANOVA showed a significant difference in stability-gap area between model-size conditions, but no significant effect of task order or interaction. Descriptively, Pythia-6.9B showed the smallest gaps across both task orders, although this is confounded by model-specific learning rates. LoRA experiments significantly reduced stability-gap area for Pythia-410M and Pythia-6.9B, but not for Pythia-14M. Overall, the findings suggest that stability gaps are relevant during sequential adaptation and that model size, optimization choices, and adaptation method can affect their severity.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Ven, G.M. van de and Ziser, Y. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 09 Jul 2026 12:45 |
| Last Modified: | 09 Jul 2026 12:45 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37409 |
Actions (login required)
 |
View Item |