CorrSynth -- A Correlated Sampling Method for Diverse Dataset Generation from LLMs

TL;DR

CorrSynth addresses the diversity and fidelity gaps in LLM-generated data for text classification by introducing correlated sampling across parallel generations. It extends classifier-free guidance to a parallel, multi-sequence setting using a denominator constructed from contrasting prompts, enabling $K$-class synthesis with significantly reduced forward passes. The method generalizes to $M$-CorrSynth with tunable cross- and intra-label guidance via $\gamma_{m,n}$, and includes a plausibility constraint with threshold $\alpha$ to balance diversity and coherence. Empirical results across four datasets show substantial gains in intrinsic diversity metrics and downstream DistilBERT accuracy compared to CFG and FewGen baselines, highlighting CorrSynth as a practical tool for producing diverse, faithful labeled data for downstream models.

Abstract

Large language models (LLMs) have demonstrated remarkable performance in diverse tasks using zero-shot and few-shot prompting. Even though their capabilities of data synthesis have been studied well in recent years, the generated data suffers from a lack of diversity, less adherence to the prompt, and potential biases that creep into the data from the generator model. In this work, we tackle the challenge of generating datasets with high diversity, upon which a student model is trained for downstream tasks. Taking the route of decoding-time guidance-based approaches, we propose CorrSynth, which generates data that is more diverse and faithful to the input prompt using a correlated sampling strategy. Further, our method overcomes the complexity drawbacks of some other guidance-based techniques like classifier-based guidance. With extensive experiments, we show the effectiveness of our approach and substantiate our claims. In particular, we perform intrinsic evaluation to show the improvements in diversity. Our experiments show that CorrSynth improves both student metrics and intrinsic metrics upon competitive baselines across four datasets, showing the innate advantage of our method.

Figures (5)

• Figure 1: CorrSynth introduces anti-correlation between examples, compared to few-shot generation.
• Figure 2: Generation progression from CFG and CorrSynth. We sample five generations using 3-shot prompts from IMDb. The colored lines represent the absolute difference between logits of the current generation and contrast for each generation timestep (taken as an exponential moving average).
• Figure 3: Heatmaps for label-wise cosine similarities on ToI Headlines (with Phi-3-mini) as we increase Intra-label contrast vs increasing cross-label contrast. Note that "Cross" and "Intra" in figure titles correspond to $\gamma_{cross}$ and $\gamma_{intra}$ respectively. FewGen heatmaps are provided for reference.
• Figure 4: Visualising two-dimensional text representations of generations (on ToI Headlines with Phi-3 Mini) using CorrSynth-Intra. We gradually increase guidance delta, $\delta$ in $(0.0, 0.5, 0.9)$. FewGen plot is provided as a reference to the unmodified clusters (it is equivalent to $\delta=1$ i.e. no contrast).
• Figure 5: Datamaps for DistilBERT training run on $2$K examples of ToI Headlines generated using CorrSynth-Intra using Phi-3-mini. FewGen datamap is provided for reference.