The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation

• URL: http://arxiv.org/abs/2412.04318v1
• Date: Thu, 05 Dec 2024 16:34:20 GMT
• Title: The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation
• Authors: Fredrik Carlsson, Fangyu Liu, Daniel Ward, Murathan Kurfali, Joakim Nivre,
• Abstract summary: This paper introduces the counter-intuitive generalization results of overfitting pre-trained large language models on very small datasets. We find that by further fine-tuning these models to achieve a near-zero training loss on a small set of samples, the long-sequence generative capabilities are greatly enhanced.
• Score: 15.904856111636851
• License:
• Abstract: This paper introduces the counter-intuitive generalization results of overfitting pre-trained large language models (LLMs) on very small datasets. In the setting of open-ended text generation, it is well-documented that LLMs tend to generate repetitive and dull sequences, a phenomenon that is especially apparent when generating using greedy decoding. This issue persists even with state-of-the-art LLMs containing billions of parameters, trained via next-token prediction on large datasets. We find that by further fine-tuning these models to achieve a near-zero training loss on a small set of samples -- a process we refer to as hyperfitting -- the long-sequence generative capabilities are greatly enhanced. Greedy decoding with these Hyperfitted models even outperform Top-P sampling over long-sequences, both in terms of diversity and human preferences. This phenomenon extends to LLMs of various sizes, different domains, and even autoregressive image generation. We further find this phenomena to be distinctly different from that of Grokking and double descent. Surprisingly, our experiments indicate that hyperfitted models rarely fall into repeating sequences they were trained on, and even explicitly blocking these sequences results in high-quality output. All hyperfitted models produce extremely low-entropy predictions, often allocating nearly all probability to a single token.

Related papers

• On the Statistical Capacity of Deep Generative Models [10.288413514555861]
  We show that deep generative models can only generate concentrated samples that exhibit light tails. These results shed light on the limited capacity of common deep generative models to handle heavy tails.
  arXiv  Detail & Related papers  (2025-01-14T00:39:46Z)
• Non-autoregressive Sequence-to-Sequence Vision-Language Models [63.77614880533488]
  We propose a parallel decoding sequence-to-sequence vision-language model that marginalizes over multiple inference paths in the decoder. The model achieves performance on-par with its state-of-the-art autoregressive counterpart, but is faster at inference time.
  arXiv  Detail & Related papers  (2024-03-04T17:34:59Z)
• Timer: Generative Pre-trained Transformers Are Large Time Series Models [83.03091523806668]
  This paper aims at the early development of large time series models (LTSM) During pre-training, we curate large-scale datasets with up to 1 billion time points. To meet diverse application needs, we convert forecasting, imputation, and anomaly detection of time series into a unified generative task.
  arXiv  Detail & Related papers  (2024-02-04T06:55:55Z)
• Amortizing intractable inference in large language models [56.92471123778389]
  We use amortized Bayesian inference to sample from intractable posterior distributions. We empirically demonstrate that this distribution-matching paradigm of LLM fine-tuning can serve as an effective alternative to maximum-likelihood training. As an important application, we interpret chain-of-thought reasoning as a latent variable modeling problem.
  arXiv  Detail & Related papers  (2023-10-06T16:36:08Z)
• Mitigating the Learning Bias towards Repetition by Self-Contrastive Training for Open-Ended Generation [92.42032403795879]
  We show that pretrained language models (LMs) such as GPT2 still tend to generate repetitive texts. We attribute their overestimation of token-level repetition probabilities to the learning bias. We find that LMs use longer-range dependencies to predict repetitive tokens than non-repetitive ones, which may be the cause of sentence-level repetition loops.
  arXiv  Detail & Related papers  (2023-07-04T07:53:55Z)
• SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking [60.109453252858806]
  A maximum-likelihood (MLE) objective does not match a downstream use-case of autoregressively generating high-quality sequences. We formulate sequence generation as an imitation learning (IL) problem. This allows us to minimize a variety of divergences between the distribution of sequences generated by an autoregressive model and sequences from a dataset. Our resulting method, SequenceMatch, can be implemented without adversarial training or architectural changes.
  arXiv  Detail & Related papers  (2023-06-08T17:59:58Z)
• A Simple Explanation for the Phase Transition in Large Language Models with List Decoding [3.898689841227059]
  We show that large language models (LLM) exhibit emergent abilities that are not present in small models. We use a list decoder that keeps a list of candidate sequences at each step and defers the generation of the output sequence at the end.
  arXiv  Detail & Related papers  (2023-03-23T09:00:07Z)
• Mutual Exclusivity Training and Primitive Augmentation to Induce Compositionality [84.94877848357896]
  Recent datasets expose the lack of the systematic generalization ability in standard sequence-to-sequence models. We analyze this behavior of seq2seq models and identify two contributing factors: a lack of mutual exclusivity bias and the tendency to memorize whole examples. We show substantial empirical improvements using standard sequence-to-sequence models on two widely-used compositionality datasets.
  arXiv  Detail & Related papers  (2022-11-28T17:36:41Z)
• Deep Generative model with Hierarchical Latent Factors for Time Series Anomaly Detection [40.21502451136054]
  This work presents DGHL, a new family of generative models for time series anomaly detection. A top-down Convolution Network maps a novel hierarchical latent space to time series windows, exploiting temporal dynamics to encode information efficiently. Our method outperformed current state-of-the-art models on four popular benchmark datasets.
  arXiv  Detail & Related papers  (2022-02-15T17:19:44Z)
• Bayesian Inference in High-Dimensional Time-Serieswith the Orthogonal Stochastic Linear Mixing Model [2.7909426811685893]
  Many modern time-series datasets contain large numbers of output response variables sampled for prolonged periods of time. In this paper, we propose a new Markov chain Monte Carlo framework for the analysis of diverse, large-scale time-series datasets.
  arXiv  Detail & Related papers  (2021-06-25T01:12:54Z)

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.