CodeAxiom
AVIXOSEC LABS CODEAXIOM PUBLIC PLAN

Verifier-guided coding model training

Train coding models through executable verification.

CodeAxiom is a public training plan and verifier demo for a future specialized coding model. It shows the benchmark targets, training stack, data policy, and compute path before any model claims are made.

Open repo Read brief Email
01

Code that runs, not code that sounds right.

The planned model learns from compiler results, unit tests, patch application, runtime errors, repository checks, and benchmark harnesses.

02

Targeting real coding evaluations

The benchmark plan covers LiveCodeBench, HumanEval+, MBPP+, MultiPL-E, Aider editing, RepoQA, SWE-bench Verified, and long-horizon SWE-style tasks.

03

No benchmark claims before evaluation

The repo has a local verifier demo and a training plan. Public scores come only after the checkpoint, harness, compute, and contamination notes are recorded.

04

Training stack

The planned path is baseline evaluation, core SFT, edit SFT, multilingual compiler loops, long-repo training, patch search, CodeWorldModel, verifier RL, and one-shot distillation.

benchHumanEval+, MBPP+, LiveCodeBench
repoAider editing, RepoQA, SWE-bench Verified
trainSFT, edit training, compiler feedback, patch search
computecompute plan, measured reports, contamination notes
05

Roadmap: from verifier demo to trained model.

The first public step is the local verifier demo. The next steps are dataset review, baseline evaluation, small checkpoints, edit training, repository tasks, and measured benchmark reports.

Compute support is described in the funding brief, but the site focuses on the project: training code models with execution feedback and publishing measured results.

contact@avixosec.xyz