@inproceedings{schoch-etal-2026-pram,
title = "{P}r{AM}: Offline Checkpoint Selection for Post-Hoc Training Data Attribution",
author = "Schoch, Stephanie and
Pathak, Sidhya and
Ji, Yangfeng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1521/",
pages = "32387--32400",
ISBN = "979-8-89176-462-0",
abstract = "Training data attribution scores how training examples affect downstream model performance. For models released with many intermediate checkpoints, attribution can be computed along the saved training trajectory, but evaluating every checkpoint is often prohibitively expensive. We introduce PrAM (Probe-based Attribution Mass), a post-hoc offline checkpoint selector for training data attribution. PrAM runs a small attribution probe at every checkpoint, selects checkpoints with high absolute attribution mass, and runs full attribution only on that subset. Across Pythia-14M, 70M, and 160M, PrAM closely matches a full attribution-pool reference and outperforms heuristic rules based on recency, spacing, learning rate, and validation loss. Beyond rank correlation, PrAM preserves the domain-by-target attribution tables and ranked training examples produced by full saved-checkpoint attribution, enabling scalable post-hoc analyses of checkpointed model releases. We support this with application-based analyses scaling to 410M and 1B parameter models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="schoch-etal-2026-pram">
<titleInfo>
<title>PrAM: Offline Checkpoint Selection for Post-Hoc Training Data Attribution</title>
</titleInfo>
<name type="personal">
<namePart type="given">Stephanie</namePart>
<namePart type="family">Schoch</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sidhya</namePart>
<namePart type="family">Pathak</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yangfeng</namePart>
<namePart type="family">Ji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Training data attribution scores how training examples affect downstream model performance. For models released with many intermediate checkpoints, attribution can be computed along the saved training trajectory, but evaluating every checkpoint is often prohibitively expensive. We introduce PrAM (Probe-based Attribution Mass), a post-hoc offline checkpoint selector for training data attribution. PrAM runs a small attribution probe at every checkpoint, selects checkpoints with high absolute attribution mass, and runs full attribution only on that subset. Across Pythia-14M, 70M, and 160M, PrAM closely matches a full attribution-pool reference and outperforms heuristic rules based on recency, spacing, learning rate, and validation loss. Beyond rank correlation, PrAM preserves the domain-by-target attribution tables and ranked training examples produced by full saved-checkpoint attribution, enabling scalable post-hoc analyses of checkpointed model releases. We support this with application-based analyses scaling to 410M and 1B parameter models.</abstract>
<identifier type="citekey">schoch-etal-2026-pram</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1521/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>32387</start>
<end>32400</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T PrAM: Offline Checkpoint Selection for Post-Hoc Training Data Attribution
%A Schoch, Stephanie
%A Pathak, Sidhya
%A Ji, Yangfeng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F schoch-etal-2026-pram
%X Training data attribution scores how training examples affect downstream model performance. For models released with many intermediate checkpoints, attribution can be computed along the saved training trajectory, but evaluating every checkpoint is often prohibitively expensive. We introduce PrAM (Probe-based Attribution Mass), a post-hoc offline checkpoint selector for training data attribution. PrAM runs a small attribution probe at every checkpoint, selects checkpoints with high absolute attribution mass, and runs full attribution only on that subset. Across Pythia-14M, 70M, and 160M, PrAM closely matches a full attribution-pool reference and outperforms heuristic rules based on recency, spacing, learning rate, and validation loss. Beyond rank correlation, PrAM preserves the domain-by-target attribution tables and ranked training examples produced by full saved-checkpoint attribution, enabling scalable post-hoc analyses of checkpointed model releases. We support this with application-based analyses scaling to 410M and 1B parameter models.
%U https://aclanthology.org/2026.findings-emnlp.1521/
%P 32387-32400
Markdown (Informal)
[PrAM: Offline Checkpoint Selection for Post-Hoc Training Data Attribution](https://aclanthology.org/2026.findings-emnlp.1521/) (Schoch et al., Findings 2026)
ACL