CodeAxiom
AVIXOSEC LABS CODEAXIOM AXIOMCODE

Executable verification for coding models

CodeAxiom builds AxiomCode.

CodeAxiom builds AxiomCode, a coding model that turns programming requests into working, verified code changes. The first target is AxiomCode-3B, based on Qwen2.5-Coder-3B.

Open repo Read brief Email
01

Base model vs trained model, same tests.

The pipeline compares Qwen2.5-Coder-3B base and AxiomCode-3B on the same verified programming tasks with the same runner.

02

AxiomRepair-20

The first suite records patch apply rate, pass@1, test pass rate, runtime, and failure category. Results are TBD until the runs exist.

03

AxiomTrace data

Training examples come from verified coding-task traces: task, files, command output, unified diff, verification result, and split metadata.

04

Measured before and after.

The first report compares Qwen2.5-Coder-3B base against AxiomCode-3B. No score is published before the benchmark runs.

baseQwen2.5-Coder-3B
modelAxiomCode-3B
suiteAxiomRepair-20
reportbase vs trained, TBD results
05

Modal is the compute backend.

The repo is set up for local smoke checks first, then Modal import checks, serving, base evaluation, LoRA training, trained evaluation, and compare reports.

Compute support helps scale the first measured training run. The public identity stays the same: CodeAxiom builds AxiomCode through executable verification.

contact@avixosec.xyz