ResoFilter: Fine-grained Synthetic Data Filtering for Large Language Models through Data-Parameter Resonance Analysis

Zeao Tu, Xiangdi Meng, Yu He, Zihan Yao, Tianyu Qi, Jun Liu, Ming LiNAACL 2025 · DOI

Abstract

ResoFilter is a novel data-filtering framework designed to identify high-quality synthetic training samples by measuring their impact on a model's internal weights during a single fine-tuning step. By filtering out "noisy" data that causes excessive parameter disruption—a phenomenon termed "data-parameter resonance"—researchers achieved performance matching or exceeding full-scale fine-tuning using only 50% of the original dataset. This approach offers a model-centric alternative to traditional output-based metrics like loss or perplexity.

Introduction

As Large Language Model (LLM) training scales, the reliance on synthetic data generated via techniques like Self-Instruct has increased. However, simply increasing data volume often leads to diminishing returns due to the introduction of low-quality or "noisy" samples. Existing filtering methods typically rely on expensive "teacher" models (e.g., GPT-4) which inherit specific biases, or surface-level metrics like loss and perplexity that fail to capture how data alters a model's internal knowledge structure.

ResoFilter addresses these limitations by shifting the focus from model outputs to internal parameter dynamics, identifying data that "resonates" constructively with the model's existing weights.

Methodology: Data-Parameter Resonance Analysis

The core of ResoFilter is the measurement of weight changes ($\Delta W$) induced by individual data points. The process treats the LLM as a physical system where high-quality information causes stable, incremental updates, while low-quality data causes volatile, disruptive shifts.

The ResoFilter workflow: (1) Calculating weight changes via one-shot fine-tuning and (2) Sorting and filtering the dataset based on the resulting 'diff' scores.

The ResoFilter Workflow

  1. Single-Shot Fine-Tuning: For each sample in a synthetic dataset, a single forward and backward propagation pass is performed on a base model ($M_0$).
  2. Weight Difference Calculation: The difference in weights between the original and updated model is calculated. Ablation studies identified the $W_{up}$ matrix (the up-projection layer in MLP modules) within the last $n$ transformer layers as the most sensitive indicator of data quality.
  3. Scoring and Filtering: Samples are assigned a "characteristic score" based on these weight changes. Data causing smaller, more stable differences is prioritized, while samples causing massive shifts are discarded as "dirty."
  4. Final Training: The model undergoes standard fine-tuning on the curated high-quality subset.

Experimental Results

The researchers evaluated ResoFilter across mathematical (MetaMath), coding (evol-codealpaca), and general domains (DOLLY, Open Assistant). Key findings include:

  • Efficiency: In mathematical tasks (GSM8k), ResoFilter matched full-dataset performance using only 50% of the data.
  • Cross-Model Scalability: Data filtered using a small model (e.g., Gemma2-2B) remained high-quality for significantly larger models (e.g., Llama2-70B), which saw a 3.2% relative improvement over random sampling. This suggests that data quality is an intrinsic property detectable across architectures.
  • Layer Sensitivity: Semantic changes driving performance were found to be concentrated in deeper layers (layers 20-26), whereas shallow layers (1-8) provided poor quality indicators.

Discussion

Analysis of rejected "dirty" data revealed patterns of shorter lengths, rare vocabulary, and low internal consistency, suggesting these samples represent outliers that confuse the model. While ResoFilter requires an initial pass for every data point, its ability to identify high-quality data without relying on proprietary teacher models makes it a scalable and less biased alternative for synthetic data curation.

Conclusion

ResoFilter demonstrates that "listening" to a model's internal resonance provides a superior signal for data quality than external probability metrics. By focusing on parameter stability, ResoFilter enables more efficient LLM training, proving that in the era of synthetic data, precision is more valuable than sheer volume.

Summary updated

BibTeX
@inproceedings{tu2025resofilter,
  author = {Zeao Tu and Xiangdi Meng and Yu He and Zihan Yao and Tianyu Qi and Jun Liu and Ming Li},
  title = {{ResoFilter: Fine-grained Synthetic Data Filtering for Large Language Models through Data-Parameter Resonance Analysis}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.299},
}
Figure