Clear Preferences Leave Traces: Reference Model-Guided Sampling for Preference Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Diwan, Nirav, Ergen, Tolga, Shim, Dongsub, Lee, Honglak
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909466352418816
author Diwan, Nirav
Ergen, Tolga
Shim, Dongsub
Lee, Honglak
author_facet Diwan, Nirav
Ergen, Tolga
Shim, Dongsub
Lee, Honglak
contents Direct Preference Optimization (DPO) has emerged as a de-facto approach for aligning language models with human preferences. Recent work has shown DPO's effectiveness relies on training data quality. In particular, clear quality differences between preferred and rejected responses enhance learning performance. Current methods for identifying and obtaining such high-quality samples demand additional resources or external models. We discover that reference model probability space naturally detects high-quality training samples. Using this insight, we present a sampling strategy that achieves consistent improvements (+0.1 to +0.4) on MT-Bench while using less than half (30-50%) of the training data. We observe substantial improvements (+0.4 to +0.98) for technical tasks (coding, math, and reasoning) across multiple models and hyperparameter settings.
format Preprint
id arxiv_https___arxiv_org_abs_2501_15109
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Clear Preferences Leave Traces: Reference Model-Guided Sampling for Preference Learning
Diwan, Nirav
Ergen, Tolga
Shim, Dongsub
Lee, Honglak
Machine Learning
Artificial Intelligence
Direct Preference Optimization (DPO) has emerged as a de-facto approach for aligning language models with human preferences. Recent work has shown DPO's effectiveness relies on training data quality. In particular, clear quality differences between preferred and rejected responses enhance learning performance. Current methods for identifying and obtaining such high-quality samples demand additional resources or external models. We discover that reference model probability space naturally detects high-quality training samples. Using this insight, we present a sampling strategy that achieves consistent improvements (+0.1 to +0.4) on MT-Bench while using less than half (30-50%) of the training data. We observe substantial improvements (+0.4 to +0.98) for technical tasks (coding, math, and reasoning) across multiple models and hyperparameter settings.
title Clear Preferences Leave Traces: Reference Model-Guided Sampling for Preference Learning
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2501.15109