Measuring the paper's footprint in language-model training

The question: how much of what language models learned about Aβ*56 came from the retracted paper and the papers built on it, and do models know that it fell? Three measurements, from text to model.

The next stage is described in unlearning_plan.md: claim-level adjudication, a completed context-control pilot on gpt-oss:20b, and a proposed comparison of correction fine-tuning with approximate unlearning. Citation reach and open-text availability are candidate exposure measures, not proof of causal influence on a particular model.

1. Which of the citing papers are in open training corpora

The 1,240 papers citing the retracted paper (first step) and the ~45,000 citing one of those (second step) identify text in which its claims may have propagated. Their citation links do not establish that a particular model learned those claims. Three public corpora can be checked directly.

Corpus	What it holds	How to test membership
The Pile (Gao et al., 2020, arXiv:2101.00027; datasheet Biderman et al., 2022, arXiv:2201.07311)	PubMed Central: the PMC Open Access subset, about 90 GiB, 14.4% of training weight at 2 epochs. PubMed Abstracts: abstracts of about 30 million PubMed records (19 GiB, 3.07%), excluding articles already in PMC and about a third dropped for malformed titles or abstracts	Current abstract availability and PMC OA membership are eligibility proxies. Exact inclusion requires checking the historical corpus snapshot, dates, exclusions and preprocessing; these flags alone do not establish membership.
peS2o (Soldaini & Lo, Hugging Face allenai/peS2o) and Dolma (Soldaini et al., 2024, arXiv:2402.00159), which contains it	About 40 million open-access papers from S2ORC; v1 knowledge cutoff 3 January 2023	Semantic Scholar's open-access flag on each citing paper approximates it; exact membership needs the dataset's corpus IDs
Models trained on them	GPT-Neo, GPT-J, GPT-NeoX, Pythia (The Pile); OLMo (Dolma); and, less transparently, most models whose cards list PubMed or PMC	Compare by training cutoff rather than by corpus

scripts/corpus_membership.py converts PMIDs to PMCIDs through E-utilities, asks PMC which are in the Open Access subset, and writes data/corpus_membership.csv. The second-step flags from scripts/fetch_hop2.py already give how many of the ~45,000 are in PMC at all.

2. How often the claim and its correction appear in training text

Verified document membership would establish inclusion in a particular dataset, not exposure by every model. The current availability flags do not establish historical membership. Two complementary counts describe the accessible text, without establishing a model's causal training exposure.

• String counts in whole corpora. infini-gram (Liu et al., 2024, arXiv:2401.17377; API at infini-gram.readthedocs.io) counts any string in Dolma v1.7 (v4_dolma-v1_7_llama) and The Pile's training set (v4_piletrain_llama). Count the claim strings (Aβ*56, Abeta*56, Aβ star 56, 56-kDa, Aβ*56 impairs memory) against the correction strings (Lesné near retract, Aβ*56 near artifact, manipulated). The ratio is the corpus's own balance between the claim and its doubt. scripts/corpus_membership.py --strings does this.
• Citing sentences by period. pubmed-stream's reading of the citing papers' full text, frozen in data/citing_papers.csv, says how each paper used the retracted one. By period, among the papers that could be read:

Period	Names Aβ*56	Cites it in general	Discusses the doubts or retraction
2006–2011	71	220	1
2012–2016	59	253	0
2017–2021	42	170	0
2022–2023	7	33	4
2024–2026	1	11	7

Until 2022, essentially every sentence a model could have read that cited the paper either took Aβ*56 as a finding or used the paper as support for a general claim; the corrective sentences are 12 of 879 and all but one are from 2022 on. Full text for 1,156 of the 1,240 citing papers is in PubMed Central, and 301 of them are reviews, the papers most likely to be paraphrased into textbooks and other secondary text.

3. What models say

probes/questions.yaml holds the questions, grouped by the four classes in docs/what_was_wrong.md (Aβ*56-specific claims, general oligomer claims, over-correction traps, retraction awareness), each with the points a correct answer must contain and the errors to score. scripts/probe_models.py asks each question of:

• local models through Ollama, with tools off by construction (gpt-oss:20b, qwen3.5, gemma4 are installed; add models with cutoffs before July 2022, such as Llama 2, and between July 2022 and June 2024, such as Llama 3);
• the installed assistants through pubmed-stream's ask_assistant, which runs Claude Code, Codex or Gemini CLI in an empty folder with their tools turned off, so the answer is parametric knowledge only.

Each question is asked three times per model; the answers are saved as JSON lines under probes/answers/ and scored by hand with the rubric in the questions file (an LLM judge can pre-score, but the final scores are read by a person). The analysis crosses the score with the model's published training cutoff relative to the three dates: 21 July 2022, 24 June 2024 and 23 December 2025.

Prior work found the same failure for other retracted papers: ChatGPT answered questions from retracted cancer-imaging papers without noting the retraction in most cases (Li & Wang, 2025, 10.1016/j.jare.2025.03.020); across 217 retracted or otherwise concerning studies, none of 6,510 ChatGPT reports mentioned a retraction (Thelwall et al., 2025, 10.1002/leap.2018); no AI tool handled 15 retracted articles consistently (Labenbacher et al., 2026, 10.2196/88766); and models' awareness of retractions is patchy (Thelwall, 2026, 10.1108/lm-04-2026-0052). Adetayo (2026, 10.1177/02663821261460796) argues the mechanism: publisher licensing deals deliver whole archives into training data without retraction filtering. This study adds what those did not have: one paper with a measured citation tree, a classified reading of how each citing paper used it, and a claim with a known correct status in three parts.

String counts, measured 20 October 2026

infini-gram counts of each string in Dolma v1.7 (the corpus of the OLMo models, assembled before the retraction) and in The Pile's training set (assembled in 2020). The claim strings are present; the correction strings are absent, except the title of the Science investigation, which appears 12 times in Dolma.

String	Kind	Dolma v1.7	Pile (train)
Aβ*56	claim	1,960	85
Abeta*56	claim	80	15
Aβ*56 impairs memory	claim	7	0
56-kDa soluble	claim	26	10
Lesné	name	2,477	272
Blots on a field	correction	12	0
Lesné retracted, Aβ*56 retracted, Aβ*56 artifact, Lesné image manipulation, Aβ*56 could not be replicated	correction	0	0

Counts are of exact token sequences under the Llama-2 tokenizer, so paraphrases of the correction are not counted; the asymmetry is still the point. data/corpus_strings.json holds the raw result.

4. Controls

• A retracted paper with an older, louder correction: Wakefield et al. 1998 (PMID 9500320), retracted in 2010, is pubmed-stream's own example in its README and lets the same pipeline run on a case where the correction dominates the corpus.
• A non-retracted paper of similar age and citation count from the same field, for the baseline rate at which models hedge.

Status

Step	Script	State
Freeze the citing papers from tutorial 8's saved graph	scripts/freeze_from_kg.py	Done, snapshot of 19 October 2026
Fetch the second step and size the field	scripts/fetch_hop2.py	See data/reach.json
String counts in Dolma and The Pile	scripts/corpus_membership.py --strings	Done, 20 October 2026
PMC Open Access membership of the citing papers	scripts/corpus_membership.py --membership	Done, 20 October 2026: 535 of 1,240 and 29,875 of 44,957 in the Open Access subset
Model probes	scripts/probe_models.py, probes/questions.yaml	Runner tested on gpt-oss:20b (two questions: both answered with confident fabrications and no retraction); full run not started
Analysis notebooks and figures	notebooks/	Not started