UniGeM : Uni fying Data Mixing and Selection via G eometric E xploration and M ining
Changhao Wang, Yunfei Yu, Xinhao Yao, Jiaolong Yang, Riccardo Cantoro, Chaobo Li, Qing Cui, Jun Zhou · 2026
Read UniGeM : Uni fying Data Mixing and Selection via G eometric E xploration and M ining by Changhao Wang, Yunfei Yu, Xinhao Yao, Jiaolong Yang, Riccardo Cantoro, Chaobo Li, Qing Cui, Jun Zhou (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03772.txt.
Description
Pretraining data curation usually mixes sources in one step and filters samples in another, and the split can break structure inside code corpora. Changhao Wang and 7 co-authors unify the two steps in UniGeM, presented in a 2026 paper. The method treats curation as approximating a data manifold, with no proxy models or external reference datasets required.
A macro-exploration stage learns mixing weights through stability-based clustering. A micro-mining stage then filters individual instances by their geometric distribution, preserving logical consistency. The team validates the approach by training mixture-of-experts models of eight and sixteen billion parameters on a hundred billion tokens. UniGeM doubles data efficiency over a random baseline and improves reasoning-heavy and multilingual results over state-of-the-art methods.
Catalog record
| CREATOR | Changhao Wang, Yunfei Yu, Xinhao Yao, Jiaolong Yang, Riccardo Cantoro, Chaobo Li, Qing Cui, Jun Zhou |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-03772