Research map / Large language models / Language, safety and interpretability
Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IMP-Loss and DIMP-Loss, efficient importance-weighted losses that reweight LLM-generated training data using a quality checker (trained on a tiny real dataset) and a diversity checker, improving BERT classifiers over CE, Focal Loss, and meta-learning reweighting baselines.
Problem
Models trained on LLM-generated synthetic data can underperform models trained on small real datasets because the synthetic distribution deviates from the real-world distribution; existing reweighting methods based on meta-learning are computationally expensive.
Approach
Weight each synthetic example by an importance ratio P'(y|x)/Q(y|x), where P' (quality checker) is fit on a small real dataset and Q (diversity checker) is fit on the synthetic data; DIMP-Loss replaces Q with the current model's prediction so only one extra forward pass is needed and no meta-learning is required.
Key results
- On GPT-3.5-generated Financial data, DIMP-Loss reaches 82.67% accuracy vs. 77.39% for CE-Loss and 79.46% for GPT-3.5 few-shot.
- On Tweet Irony with GPT-3.5 generated data, IMP-Loss achieves 81.89% accuracy vs. 76.91% for CE-Loss.
- DIMP-Loss total training time was about 371s vs. 2680s for SunGen and 333s for CE-Loss on the Financial benchmark.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IMP-Loss instead of CE-Loss on GPT-3.5-generated data | BERT-base classifier · Financial Phrasebank test | accuracy: 77.39 -> 82.09 | Helped |
| Use DIMP-Loss instead of CE-Loss on GPT-3.5-generated data | BERT-base classifier · Tweet Irony test | accuracy: 76.91 -> 78.44 | Helped |
| Use IMP-Loss instead of CE-Loss on GPT-3.5-generated data | BERT-base classifier · MRPC test | accuracy: 72 -> 75.83 | Helped |
| Use DIMP-Loss instead of CE-Loss on large real-world training data | BERT-base classifier · MRPC test | accuracy: 80.92 -> 82.84 | Helped |
| Add a diversity checker to DIMP-Loss (ablation) | BERT-base classifier trained on GPT-3.5-generated data · Financial Phrasebank test | accuracy: 81.35 -> 82.67 | Helped |
| Add a diversity checker to DIMP-Loss (ablation) | BERT-base classifier trained on GPT-3.5-generated data · Financial Phrasebank test | macro F1: 77.94 -> 79.53 | Helped |
| Use IMP-Loss rather than DIMP-Loss on GPT-3.5-generated data | BERT-base classifier · Tweet Irony test | accuracy: 78.44 -> 81.89 | Helped |
| Use IMP-Loss rather than DIMP-Loss on large real-world training data | BERT-base classifier · Tweet Irony test | accuracy: 69 -> 70.15 | Helped |
| Use DIMP-Loss rather than IMP-Loss on large real-world training data | BERT-base classifier · MRPC test | accuracy: 81.33 -> 82.84 | Helped |
| Use DIMP-Loss rather than IMP-Loss on noisy Financial data | BERT-base classifier · Financial Phrasebank test | accuracy: 81.6 -> 82.59 | Helped |
Surprising findings
- Quality checker trained on small real data alone (62.5% accuracy on Tweet Irony) is much weaker than CE on generated data (76.9%), yet weighting still helps.
- Removing the diversity checker from IMP-Loss lowers Financial accuracy from 82.09% to 81.35%, showing diversity weighting matters.
Limitations
- Evaluated only with BERT-base/large encoders and GPT-3.5 as generator; no evaluation with other generators or larger classifiers.
- Key assumption that the LLM approximates the real input distribution Q(x) ≈ P(x) is not verified.
- IMP-Loss requires roughly 2.2x CE training time; approximations (P' for the numerator, lower-bound derivation) are heuristic.
Experimental setup
Datasets: Financial Phrasebank Tweet Irony MRPC Tweet Sentiment Extraction
Benchmarks: Financial Phrasebank Tweet Irony MRPC
Metrics: accuracy macro F1
Baselines: CE-Loss Focal-Loss SunGen meta-learning reweighting method (cited work) GPT-3.5 few-shot
Builds on: Focal Loss SunGen online batch selection importance sampling ZeroGen
Closest papers by meaning
- Idiosyncrasies in Large Language Models (ICML 2025 · similarity 0.88)
- Escaping Collapse: The Strength of Weak Data for Large Language Model Training (NeurIPS 2025 · similarity 0.87)
- Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective (ICLR 2025 · similarity 0.87)
- Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining (ICLR 2025 · similarity 0.87)
- Synthetic Text Generation for Training Large Language Models via Gradient Matching (ICML 2025 · similarity 0.87)
- LLM Data Selection and Utilization via Dynamic Bi-level Optimization (ICML 2025 · similarity 0.87)
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.