gpt2-memorization-1-to-7-pct-exact-duplicates
IN premise — entries/2026/06/21/wiki-Large_language_model-chunk-3.md
Created 2026-06-21T09:50:09+00:00
GPT-2 models were measured to contain 1% to approximately 7% exact duplicate outputs from training data, and repetition prompting can trigger memorized excerpts.
Summary
GPT-2 doesn't just generate novel text; a meaningful slice of its output, somewhere between 1 and 7 percent, is word-for-word copied from the material it was trained on, and this isn't hidden, since simple prompts asking it to repeat can deliberately pull out those memorized passages. That matters because it means the model can leak training data verbatim, which raises real concerns about data ownership, privacy, and whether its outputs should be treated as genuinely original.
Dependents
These beliefs depend on this one:
- IN gpt2-foreshadowed-capability-risk-dual-scaling — GPT-2's staged release over misuse concerns, combined with measurements showing 1-7% exact duplicate training data in its outputs, illustrated early tensions between scaling language models and managing associated risks such as memorization and potential misuse.
- IN memorization-creates-attack-surface-for-training-poisoning — LLM training data memorization (measured at 1-7% exact duplicates in GPT-2) is a plausible mechanism through which deliberate data poisoning (grooming) could materialize in model outputs — if poisoned content is memorized, it may be retrievable at inference time, suggesting that memorization rates could serve as one indicator of susceptibility to poisoning attacks.