Fishing For Magikarp Automatically Detecting Under Trained Tokens In Large Language Models 2024 05 08
Source
cohere.com
Author
Cohere
Date
Context
A LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition →'s vocabulary can contain text units that scarcely appeared during its training. The model can therefore receive a valid input for which it learned little useful behavior. The research abstract published by Cohere identifies this mismatch as the source of under-trained, sometimes disruptive tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →.
Fishing for Magikarp reports automated detection methods combining vocabulary analysis, signals in model weights and prompting tests. The authors found these tokens across multiple models. That makes them a systematic testing target, although the abstract does not establish a universal repair or quantify reliability improvements.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
postCohere releases open-weights translation model North Small Translate
articleThe Data Provenance Initiative A Large Scale Audit Of Dataset Licensing And Attribution In Ai 2023 10 25
articleFishing For Magikarp Automatically Detecting Under Trained Tokens In Large Language Models 2024 09 24Cohere
articleBreaking Down Model Vocabulary Barriers With Tokenizer TransplantationArcee AI
articleOne Tokenizer To Rule Them All Emergent Language Plasticity Via Multilingual Tokenizers 2025 05 30Cohere