Research map / Label-efficient and robust learning / Domain adaptation and generalization
Better Fine-Tuning via Instance Weighting for Text Classification
Zhi Wang, Wei Bi, Yan Wang, Xiaojiang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add soft-mixing instance weighting to fine-tuning | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 52.46 -> 53.54 | Improved |
| add soft-mixing instance weighting to fine-tuning | IW-Fit fine-tuning · Yelp-2015 to Yelp-2016 | accuracy (%): 57.18 -> 58.30 | Improved |
| add soft-mixing instance weighting to fine-tuning | IW-Fit fine-tuning · Yelp-2015 to Yelp-2017 | accuracy (%): 55.08 -> 55.90 | Improved |
| use prediction-loss (gradient) weights instead of uniform (ablation) | IW-Fit fine-tuning · Yelp-2015 to Yelp-2017 | accuracy (%): 55.08 -> 55.08 | No clear effect |
| use variance-of-loss weights instead of prediction-loss weights (ablation) | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 52.80 -> 53.20 | Improved |
| use variance-of-loss weights instead of prediction-loss weights (ablation) | IW-Fit fine-tuning · Yelp-2015 to Yelp-2016 | accuracy (%): 57.82 -> 57.60 | Hurt |
| soft-mixing instead of hard-mixing of the two metrics (ablation) | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 53.44 -> 53.54 | Improved |
| fine-tune over pre-training baseline: fine-tuned Uniform vs All (joint training) (ablation) | Fine-tuned Uniform · Amazon to Yelp | accuracy (%): 50.94 -> 52.46 | Improved |
| replace LSTM with Text-CNN encoder, soft-mixing vs uniform (ablation) | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 47.94 -> 48.56 | Improved |
| fix encoder (fine-tune only FC) vs fine-tune all, Soft-mixing (ablation) | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 53.54 -> 48.23 | Hurt |
| fix word embedding only, Soft-mixing (ablation) | IW-Fit fine-tuning · Yelp-2015 to Yelp-2016 | accuracy (%): 58.30 -> 57.60 | Hurt |
| vary target fine-tuning data to 25%, soft-mixing vs uniform (ablation) | IW-Fit fine-tuning · Amazon to Yelp | accuracy (%): 49.72 -> 50.46 | Improved |
Closest papers by meaning
- LIFT: Language-Interfaced Fine-Tuning for Non-language Machine Learning Tasks (NeurIPS 2022 · similarity 0.84)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.84)
- Pay Attention to Small Weights (NeurIPS 2025 · similarity 0.83)
- Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification (ICLR 2025 · similarity 0.82)
- Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining (ICLR 2025 · similarity 0.82)
- Chameleon: A Flexible Data-mixing Framework for Language Model Pretraining and Finetuning (ICML 2025 · similarity 0.82)
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.
