Scalable Influence and Fact Tracing for Large Language Model Pretraining

• URL: http://arxiv.org/abs/2410.17413v1
• Date: Tue, 22 Oct 2024 20:39:21 GMT
• Title: Scalable Influence and Fact Tracing for Large Language Model Pretraining
• Authors: Tyler A. Chang, Dheeraj Rajagopal, Tolga Bolukbasi, Lucas Dixon, Ian Tenney,
• Abstract summary: Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples. This paper refines existing gradient-based methods to work effectively at scale.
• Score: 14.598556308631018
• License:
• Abstract: Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual attribution and causal influence. With increasing model size and training tokens, we find that influence more closely aligns with attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names.

Related papers

• Influence Functions for Scalable Data Attribution in Diffusion Models [52.92223039302037]
  Diffusion models have led to significant advancements in generative modelling. Yet their widespread adoption poses challenges regarding data attribution and interpretability. In this paper, we aim to help address such challenges by developing an textitinfluence functions framework.
  arXiv  Detail & Related papers  (2024-10-17T17:59:02Z)
• ZZU-NLP at SIGHAN-2024 dimABSA Task: Aspect-Based Sentiment Analysis with Coarse-to-Fine In-context Learning [0.36332383102551763]
  The DimABSA task requires fine-grained sentiment intensity prediction for restaurant reviews. We propose a Coarse-to-Fine In-context Learning method based on the Baichuan2-7B model for the DimABSA task.
  arXiv  Detail & Related papers  (2024-07-22T02:54:46Z)
• The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes [30.30769701138665]
  We introduce and explore the Mirrored Influence Hypothesis, highlighting a reciprocal nature of influence between training and test data. Specifically, it suggests that evaluating the influence of training data on test predictions can be reformulated as an equivalent, yet inverse problem. We introduce a new method for estimating the influence of training data, which requires calculating gradients for specific test samples, paired with a forward pass for each training point.
  arXiv  Detail & Related papers  (2024-02-14T03:43:05Z)
• Revisiting Demonstration Selection Strategies in In-Context Learning [66.11652803887284]
  Large language models (LLMs) have shown an impressive ability to perform a wide range of tasks using in-context learning (ICL) In this work, we first revisit the factors contributing to this variance from both data and model aspects, and find that the choice of demonstration is both data- and model-dependent. We propose a data- and model-dependent demonstration selection method, textbfTopK + ConE, based on the assumption that textitthe performance of a demonstration positively correlates with its contribution to the model's understanding of the test samples.
  arXiv  Detail & Related papers  (2024-01-22T16:25:27Z)
• Influence Scores at Scale for Efficient Language Data Sampling [3.072340427031969]
  "influence scores" are used to identify important subsets of data. In this paper, we explore the applicability of influence scores in language classification tasks.
  arXiv  Detail & Related papers  (2023-11-27T20:19:22Z)
• An Additive Instance-Wise Approach to Multi-class Model Interpretation [53.87578024052922]
  Interpretable machine learning offers insights into what factors drive a certain prediction of a black-box system. Existing methods mainly focus on selecting explanatory input features, which follow either locally additive or instance-wise approaches. This work exploits the strengths of both methods and proposes a global framework for learning local explanations simultaneously for multiple target classes.
  arXiv  Detail & Related papers  (2022-07-07T06:50:27Z)
• Improving Pre-trained Language Model Fine-tuning with Noise Stability Regularization [94.4409074435894]
  We propose a novel and effective fine-tuning framework, named Layerwise Noise Stability Regularization (LNSR) Specifically, we propose to inject the standard Gaussian noise and regularize hidden representations of the fine-tuned model. We demonstrate the advantages of the proposed method over other state-of-the-art algorithms including L2-SP, Mixout and SMART.
  arXiv  Detail & Related papers  (2022-06-12T04:42:49Z)
• Tracing Knowledge in Language Models Back to the Training Data [39.02793789536856]
  We introduce a new benchmark for fact tracing: tracing language models' assertions back to the training examples that provided evidence for those predictions. We evaluate influence methods for fact tracing, using well-understood information retrieval metrics.
  arXiv  Detail & Related papers  (2022-05-23T17:34:16Z)
• Combining Feature and Instance Attribution to Detect Artifacts [62.63504976810927]
  We propose methods to facilitate identification of training data artifacts. We show that this proposed training-feature attribution approach can be used to uncover artifacts in training data. We execute a small user study to evaluate whether these methods are useful to NLP researchers in practice.
  arXiv  Detail & Related papers  (2021-07-01T09:26:13Z)
• An Empirical Comparison of Instance Attribution Methods for NLP [62.63504976810927]
  We evaluate the degree to which different potential instance attribution agree with respect to the importance of training samples. We find that simple retrieval methods yield training instances that differ from those identified via gradient-based methods.
  arXiv  Detail & Related papers  (2021-04-09T01:03:17Z)

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.