Research map / Architectures and efficiency / Pruning and efficient tuning
On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · RTE | development standard deviation: 7.9 -> 2.7 | Helped |
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · MRPC | development standard deviation: 3.8 -> 0.8 | Helped |
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · CoLA | development standard deviation: 20.9 -> 1.8 | Helped |
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · RTE | development mean: 65.3 -> 67.3 | Helped |
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · MRPC | development mean: 87.8 -> 90.3 | Helped |
| Use bias correction and train for 20 epochs instead of default fine-tuning | BERT · CoLA | development mean: 51.9 -> 62.1 | Helped |
| Remove bias correction while training for only 3 epochs (ablation) | BERT fine-tuning on RTE · RTE | performance | Hurt |
| Add bias correction (ablation) | BERT fine-tuning on RTE · RTE | fine-tuning stability | Helped |
| Train for more epochs together with bias correction (ablation) | BERT fine-tuning on RTE · RTE | performance | Helped |
| Use larger learning rates with bias correction (ablation) | BERT fine-tuning on RTE · RTE | fine-tuning stability | Helped |
| Increase training from 3 to 71 epochs on 1k samples (ablation) | BERT fine-tuning · SciTail development | standard deviation: 17.9 -> 0.9 | Helped |
| Increase training from 3 to 71 epochs on 1k samples (ablation) | BERT fine-tuning · SciTail development | mean: 69.8 -> 87.5 | Helped |
Closest papers by meaning
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.88)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.86)
- Understanding Linear Probing then Fine-tuning Language Models from NTK Perspective (NeurIPS 2024 · similarity 0.86)
- PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs (ICLR 2025 · similarity 0.86)
- TuCo: Measuring the Contribution of Fine-Tuning to Individual Responses of LLMs (ICML 2025 · similarity 0.86)
- The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.