Context-Free Synthetic Data Mitigates Forgetting

• URL: http://arxiv.org/abs/2505.13811v1
• Date: Tue, 20 May 2025 01:47:31 GMT
• Title: Context-Free Synthetic Data Mitigates Forgetting
• Authors: Parikshit Bansal, Sujay Sanghavi,
• Abstract summary: We show that augmenting a fine-tuning dataset with context-free generations mitigates forgetting.<n>We present our results for OLMo-1B for pretrained-only setting and R1-Distill-Llama-8B for the reasoning setting.
• Score: 13.825822994127943
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Fine-tuning a language model often results in a degradation of its existing performance on other tasks, due to a shift in the model parameters; this phenomenon is often referred to as (catastrophic) forgetting. We are interested in mitigating this, in settings where we only have access to the model weights but no access to its training data/recipe. A natural approach is to penalize the KL divergence between the original model and the new one. Our main realization is that a simple process - which we term context-free generation - allows for an approximate unbiased estimation of this KL divergence. We show that augmenting a fine-tuning dataset with context-free generations mitigates forgetting, in two settings: (a) preserving the zero-shot performance of pretrained-only models, and (b) preserving the reasoning performance of thinking models. We show that contextual synthetic data, and even a portion of the pretraining data, are less effective. We also investigate the effect of choices like generation temperature, data ratios etc. We present our results for OLMo-1B for pretrained-only setting and R1-Distill-Llama-8B for the reasoning setting.

Related papers

• When Models Don't Collapse: On the Consistency of Iterative MLE [34.99810116340191]
  We study model collapse for maximum likelihood estimation (MLE)<n>We establish non-asymptotic bounds showing that collapse can be avoided even as the fraction of real data vanishes.<n>We prove that some assumptions (beyond MLE consistency) are indeed necessary.
  arXiv  Detail & Related papers  (2025-05-25T08:50:46Z)
• Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting [15.251425165987987]
  Fine-tuning a pre-trained model on a downstream task often degrades its original capabilities.<n>We propose a sample weighting scheme for the fine-tuning data based on the pre-trained model's losses.<n>We empirically demonstrate the efficacy of our method on both language and vision tasks.
  arXiv  Detail & Related papers  (2025-02-05T00:49:59Z)
• Attribute-to-Delete: Machine Unlearning via Datamodel Matching [65.13151619119782]
  Machine unlearning -- efficiently removing a small "forget set" training data on a pre-divertrained machine learning model -- has recently attracted interest. Recent research shows that machine unlearning techniques do not hold up in such a challenging setting.
  arXiv  Detail & Related papers  (2024-10-30T17:20:10Z)
• On conditional diffusion models for PDE simulations [53.01911265639582]
  We study score-based diffusion models for forecasting and assimilation of sparse observations. We propose an autoregressive sampling approach that significantly improves performance in forecasting. We also propose a new training strategy for conditional score-based models that achieves stable performance over a range of history lengths.
  arXiv  Detail & Related papers  (2024-10-21T18:31:04Z)
• Understanding Fine-tuning in Approximate Unlearning: A Theoretical Perspective [39.958103832214135]
  Fine-tuning (FT) methods have become one of the fundamental approaches for approximating unlearning.<n>We present the first theoretical analysis of FT methods for machine unlearning within a linear regression framework.<n>We propose a novel Retention-Based Masking (RBM) strategy that constructs a weight saliency map based on the remaining dataset.
  arXiv  Detail & Related papers  (2024-10-04T18:01:52Z)
• Low-rank finetuning for LLMs: A fairness perspective [54.13240282850982]
  Low-rank approximation techniques have become the de facto standard for fine-tuning Large Language Models. This paper investigates the effectiveness of these methods in capturing the shift of fine-tuning datasets from the initial pre-trained data distribution. We show that low-rank fine-tuning inadvertently preserves undesirable biases and toxic behaviors.
  arXiv  Detail & Related papers  (2024-05-28T20:43:53Z)
• Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias [47.79659355705916]
  Model-induced distribution shifts (MIDS) occur as previous model outputs pollute new model training sets over generations of models. We introduce a framework that allows us to track multiple MIDS over many generations, finding that they can lead to loss in performance, fairness, and minoritized group representation. Despite these negative consequences, we identify how models might be used for positive, intentional, interventions in their data ecosystems.
  arXiv  Detail & Related papers  (2024-03-12T17:48:08Z)
• How to Learn when Data Gradually Reacts to Your Model [10.074466859579571]
  We propose a new algorithm, Stateful Performative Gradient Descent (Stateful PerfGD), for minimizing the performative loss even in the presence of these effects. Our experiments confirm that Stateful PerfGD substantially outperforms previous state-of-the-art methods.
  arXiv  Detail & Related papers  (2021-12-13T22:05:26Z)
• Contrastive Model Inversion for Data-Free Knowledge Distillation [60.08025054715192]
  We propose Contrastive Model Inversion, where the data diversity is explicitly modeled as an optimizable objective. Our main observation is that, under the constraint of the same amount of data, higher data diversity usually indicates stronger instance discrimination. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CMI achieves significantly superior performance when the generated data are used for knowledge distillation.
  arXiv  Detail & Related papers  (2021-05-18T15:13:00Z)
• Variational Bayesian Unlearning [54.26984662139516]
  We study the problem of approximately unlearning a Bayesian model from a small subset of the training data to be erased. We show that it is equivalent to minimizing an evidence upper bound which trades off between fully unlearning from erased data vs. not entirely forgetting the posterior belief. In model training with VI, only an approximate (instead of exact) posterior belief given the full data can be obtained, which makes unlearning even more challenging.
  arXiv  Detail & Related papers  (2020-10-24T11:53:00Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.