Towards Active Synthetic Data Generation for Finetuning Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kessler, Samuel, Xia, Menglin, Diaz, Daniel Madrigal, Han, Dongge, Heshemi, Helia, Rajmohan, Saravan, Ruehle, Victor, Ash, Jordan T.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910016126058496
author Kessler, Samuel
Xia, Menglin
Diaz, Daniel Madrigal
Han, Dongge
Heshemi, Helia
Rajmohan, Saravan
Ruehle, Victor
Ash, Jordan T.
author_facet Kessler, Samuel
Xia, Menglin
Diaz, Daniel Madrigal
Han, Dongge
Heshemi, Helia
Rajmohan, Saravan
Ruehle, Victor
Ash, Jordan T.
contents A common and effective means for improving language model capabilities involves finetuning a ``student'' language model's parameters on generations from a more proficient ``teacher'' model. Termed ``synthetic data'', these generations are often produced before any student finetuning, but some work has considered generating new synthetic samples as training progresses. This paper studies and advocates for the latter case, where data are generated in an iterative, closed-loop fashion that is guided by the current state of the student model. For a fixed budget of generated samples, or a budget in terms of compute spent querying a teacher, we show that this curation of finetuning data affords improved student performance over static generation. Further, while there have been several LLM-specific methods proposed that operate in this regime, we find that simple, inexpensive selection criteria from the active learning literature tend to be most performant. We validate these claims across four mathematical and logical reasoning datasets using four different small language models.
format Preprint
id arxiv_https___arxiv_org_abs_2512_00884
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Towards Active Synthetic Data Generation for Finetuning Language Models
Kessler, Samuel
Xia, Menglin
Diaz, Daniel Madrigal
Han, Dongge
Heshemi, Helia
Rajmohan, Saravan
Ruehle, Victor
Ash, Jordan T.
Machine Learning
Computation and Language
A common and effective means for improving language model capabilities involves finetuning a ``student'' language model's parameters on generations from a more proficient ``teacher'' model. Termed ``synthetic data'', these generations are often produced before any student finetuning, but some work has considered generating new synthetic samples as training progresses. This paper studies and advocates for the latter case, where data are generated in an iterative, closed-loop fashion that is guided by the current state of the student model. For a fixed budget of generated samples, or a budget in terms of compute spent querying a teacher, we show that this curation of finetuning data affords improved student performance over static generation. Further, while there have been several LLM-specific methods proposed that operate in this regime, we find that simple, inexpensive selection criteria from the active learning literature tend to be most performant. We validate these claims across four mathematical and logical reasoning datasets using four different small language models.
title Towards Active Synthetic Data Generation for Finetuning Language Models
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2512.00884