SYSTEM Cited by 1 source
regress-lm¶
regress-lm is Google DeepMind's open-source library
for training and serving Regression
Language Models (RLMs) — language models that read string
representations of system state and emit numeric targets as
decoded text. Announced alongside the 2025-07-29 Google Research
post "Simulating large systems with Regression Language Models"
and the backing paper
Performance Prediction for Large Systems via Text-to-Text
Regression.
- Repository: https://github.com/google-deepmind/regress-lm
- Purpose stated in the 2025-07-29 blog post: "We also released an open-source library for the research community to leverage for any use-case" (Source: sources/2025-07-29-google-simulating-large-systems-with-regression-language-models).
Role on the wiki¶
The wiki treats regress-lm as the open-source scaffolding
around the RLM technique — distinct from the RLM model artefact
itself and from the specific
Borg-MIPS-per-GCU application. The library is
what lets external researchers reproduce the text-to-text-
regression shape on their own workloads.
The 2025-07-29 blog post does not document the library's API, dependencies, or training scripts beyond naming its existence and linking to the GitHub repo. This page is therefore a stub that will be extended when a source covers the library's internals.
What the wiki knows¶
- Published by Google DeepMind on GitHub.
- Named in the 2025-07-29 blog post as the research-community- facing output.
- Corresponds to the RLM technique described in the post and its backing paper.
What the wiki doesn't yet know¶
- Exact API shape (tokenizer, trainer, sampler).
- Relationship to upstream frameworks (JAX / Flax / PyTorch).
- Whether Google's own Borg-deployed RLM is trained via this library or via internal tooling.
- OmniPred's (the 2024 predecessor's) library relationship.
Seen in¶
- sources/2025-07-29-google-simulating-large-systems-with-regression-language-models — announces the open-source release.