Synthetic training data boosts language model factual accuracy efficiently
It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs
Computation and LanguageMachine Learning
Summary
Current language models learn from large amounts of web text, which may not be ideal for teaching reasoning or facts. The authors created a publicly available synthetic dataset called SYNTH, based on Wikipedia, designed to improve training by combining multiple training phases into one. Models trained on SYNTH perform well compared to those trained on traditional web data, using much less training material. This approach could help create reliable language models faster and with less data.
What this means in practice
- •For language model developers: Train language models with fewer data tokens while achieving higher factual accuracy using the open-source SYNTH synthetic dataset.
- •For domain-specific ai teams: Build domain-specific language models without relying on instruction or conversational data by leveraging synthetic Wikipedia-based datasets.
Authors
Pierre-Carl Langlais, Pieter Delobelle, Yannick Detrois, Pavel Chizhov, Carlos Rosas-Hinostroza, Neil Si Smail, Benjamin Burtin, Hanna Shcharbakova, Ivan Yamshchikov, Anastasia Stasenko
Abstract
Current pre-training datasets are derived from web crawls, with all their issues, and were not designed to support mid- and post-training pipelines--for instance, they contain little explicit reasoning. Thus, many frontier labs have begun to develop their own internal datasets, starting from state-of-the-art models, to augment their pre-training data mix, eg, with reasoning traces to address cold-start problems. While demonstratively effective, none of these datasets are public, and the effect of this so-called synthetic data on knowledge and skill acquisition of language models, including small ones, remains poorly understood. We present SYNTH, the first open-source synthetic corpus derived from 58,698 Wikipedia articles that collapses pre-, mid-, and post-training into a single training stage via structured amplification of curated encyclopedic seeds. We evaluate SYNTH by training a suite of models: a 56M tiny model (Monad), 0.3B-0.6B dense models (Baguettotron), and a 13B / 1B-active MoE. At iso-compute, SYNTH outperforms filtered web data, and our models remain competitive with similarly-sized open-weight baselines. Because SYNTH is back-translated from grounded passages, SYNTH-trained models achieve high factual precision despite 10-140x fewer training tokens, with memorization targeted by the seed corpus. These results show that synthetic datasets, including our SYNTH dataset, are capable of producing competitive generalist models from a fraction of the training data, enabling rapid iteration as the frontier advances. These findings open up possibilities for both generalist models with significantly increased data efficiency, as well as domain-specific models where no instruction or conversational data is available. Finally, we publicly release our SYNTH dataset and the suite of Baguettotron models under a permissive license, thus supporting open-source language model development.