Llemma releases open language models and data specialized for mathematics

The project continued pretraining Code Llama on Proof-Pile-2, a mixture of mathematical web text, scientific papers and mathematical code.

What happened?

Llemma released 7B and 34B math-specialized models that outperformed known open base models on MATH and showed tool-use and formal-proof capabilities without task-specific fine-tuning.

Why it matters

It gave the community an open base model and corpus for studying how mathematical specialization transfers across informal, tool and formal settings.

Significance rationale: Llemma established a reproducible open math-language-model stack, including model weights, Proof-Pile-2 and training code.

Historical context

Primary evidence

Evidence & status at a glance

Significance
H2 · Field milestone
Verification
Paper released
Evidence level
E3 · Inspectable artifact
AI role
Computation only
Formal assurance
Artifact available

How should these labels be read? →

Contribution & verification

Mathematical novelty
Benchmark result
Interface
Informal, Formal, Symbolic
AI contribution
The trained models perform mathematical reasoning, tool use and formal proof generation.
Human contribution
Researchers assembled Proof-Pile-2, continued pretraining the models and released the full artifacts.
Verification
Paper released · E3 · Inspectable artifact · Artifact available
Last reviewed
2026-09-11

Technical details

Systems
llemma
Organizations
eleutherai, princeton-university
Problems
math-benchmark
Methods
continued-pretraining, math-corpus, language-model

Sources & artifacts

Verification history

  • 2023-10-16 · Paper released · E2 · Research record

    The Chronicle records this event from the cited primary or official source; the status does not imply independent Chronicle verification.

  • 2026-09-11 · Paper released · E3 · Inspectable artifact

    This corpus refresh added a directly inspectable public artifact maintained by the authors or releasing organization. E3 records artifact-level inspectability and does not imply independent verification of the underlying scientific or mathematical claim.

Corrections

No corrections recorded.