Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokens

2024/01/30 by Jiacheng Liu, Liu, Jiacheng, Sewon Min +7 · 10 voices · 40 citations
Computer Science · #Natural Language Processing Techniques #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2401.17377

Abstract

Are n-gram language models still relevant in this era of neural large language models (LLMs)? Our answer is yes, and we showcase their values in both text analysis and improving neural LLMs. This was done by modernizing n-gram LMs in two aspects. First, we train them at the same data scale as neural LLMs -- 5 trillion tokens. This is the largest n-gram LM ever built. Second, existing n-gram LMs use small n which hinders their performance; we instead allow n to be arbitrarily large, by introducing a new ∞-gram LM with backoff. Instead of pre-computing n-gram count tables (which would be very expensive), we develop an engine named infini-gram -- powered by suffix arrays -- that can compute ∞-gram (as well as n-gram with arbitrary n) probabilities with millisecond-level latency. The ∞-gram framework and infini-gram engine enable us to conduct many novel and interesting analyses of human-written and machine-generated text: we find that the ∞-gram LM has fairly high accuracy for next-token prediction (47%), and can complement neural LLMs to greatly reduce their perplexity. When analyzing machine-generated text, we also observe irregularities in the machine--∞-gram agreement level with respect to the suffix length, which indicates deficiencies in neural LLM pretraining and the positional embeddings of Transformers.

Cited by

• When Trivia Is Not Trivial: Everyday Knowledge Failures in Multilingual LLMs
• ScienceMeter: Tracking Scientific Knowledge Updates in Language Models
• Frontier AI performance across the business disciplines: a case-grounded benchmark of knowledge work and analytical reasoning
• Tracing LLM Behavior to the Training Data with Empirical Next-Token Distributions
• Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models
• Speculative Speculative Decoding
• Constructions are Revealed in Word Distributions
• OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens
• Not All Data Are Unlearned Equally
• Information-Guided Identification of Training Data Imprint in (Proprietary) Large Language Models
• Shades of Zero: Distinguishing Impossibility from Inconceivability
• Reading Without a Reader: Large Language Models Collapse Reading and Writing into a Single Entangled Code
• On Finding Inconsistencies in Documents
• Counting Clues: A Lightweight Probabilistic Baseline Can Match an LLM
• Short-Context Dominance: How Much Local Context Natural Language Actually Needs?
• Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
• TrackList: Tracing Back Query Linguistic Diversity for Head and Tail Knowledge in Open Large Language Models
• Beat the long tail: Distribution-Aware Speculative Decoding for RL Training
• ChEmREF: Evaluating Language Model Readiness for Chemical Emergency Response
• Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention
• Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs
• The Illusion of Certainty: Uncertainty Quantification for LLMs Fails under Ambiguity
• LLM generation novelty through the lens of semantic similarity
• Gaperon: A Peppered English-French Generative Language Model Suite
• Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale
• Disaggregation Reveals Hidden Training Dynamics: The Case of Agreement Attraction
• Retracing the Past: LLMs Emit Training Data When They Get Lost
• Blackbox Model Provenance via Palimpsestic Membership Inference
• AI Agents as Universal Task Solvers
• LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance
• On the Representations of Entities in Auto-regressive Large Language Models
• All Code, No Thought: Current Language Models Struggle to Reason in Ciphered Language
• SoftMatcha 2: A Fast and Soft Pattern Matcher for Trillion-Scale Corpora
• Tracing the Representation Geometry of Language Models from Pretraining to Post-training
• Death of the Novel(ty): Beyond n-Gram Novelty as a Metric for Textual Creativity
• One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning
• Coherence-driven inference for cybersecurity
• Facts Fade Fast: Evaluating Memorization of Outdated Medical Knowledge in Large Language Models
• Natural Context Drift Undermines the Natural Language Understanding of Large Language Models
• Continuously Steering LLMs Sensitivity to Contextual Knowledge with Proxy Models
• Diagnosing Memorization in Chain-of-Thought Reasoning, One Token at a Time

Discussions

• Infini-Gram: Scaling unbounded n-gram language models to a trillion tokens [hn, 142 points, 58 comments]
• random post o-the-day infinigram what if we made a language model that was just a very large trie? like, we just counted all the letter sequences and what comes after them. that's infinigram. written [bsky, 57 points, 3 comments]
• We got really big Markov chains too dw arxiv.org/abs/2401.17377 [bsky, 9 points, 1 comments]
• Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokens [bsky, 7 points, 0 comments]
• Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokens [lobsters, 5 points, 4 comments]
• Not to my knowledge, but this paper is related: arxiv.org/abs/2401.17377 The dark downside of SOTA chasing is that there is no room to try ideas that don't immediately deliver comparable performance [bsky, 2 points, 1 comments]
• Counterpoint - perhaps there is still much to be done [bsky, 1 points, 0 comments]
• If you have a model that can give you the count of any arbitrary length ngram in a dataset, what's the right search pattern to find the maximum suffix of an input that exists? I got nerd sniped into t [bsky, 1 points, 1 comments]
• Thinking way too much about making chains right now... Re-reading arxiv.org/abs/2401.17377 [bsky, 1 points, 0 comments]
• Infini-Gram: Scaling unbounded n-gram language models to a trillion tokens (arxiv.org) Main Link | Discussion [bsky, 0 points, 0 comments]

Related

• N-gram Is Back: Residual Learning of Neural Text Generation with n-gram Language Model
• Evaluating n-Gram Novelty of Language Models Using Rusty-DAWG
• A Survey of Large Language Models
• Improving N-gram Language Models with Pre-trained Deep Transformer
• Can Transformers Learn n-gram Language Models?