Hugging Face's logo

Snapkitty
/
pax-coder

custom
code
sovereign-compute
Model card Files
xet
 Community
pax-coder / PACKAGE.md
SNAPKITTYWEST's picture
SNAPKITTYWEST
chore: push pax-coder from SNAPKITTYWEST GitHub
ef6eb55 verified
preview code
|
Raw
History Blame Contribute Delete
2.98 kB

PAX-Coder Package Manifest

Package: pax-coder Version: 1.0.0 Release date: 2026-08-18 Repository: SNAPKITTYWEST/pax-coder

Package Identity

PAX-Coder is the institutional package for proof-carrying GPU kernel generation around the PAX axiom basis, Lean 4 proof modules, CUDA/PTX implementation surfaces, Futhark functional references, and training-data export.

Contents

Path Package role
README.md Institutional entry point
ABOUT.md Short project overview
LICENSE.tri License structure
VERSION Version marker
CHANGELOG.md Release history
RELEASE_NOTES.md Current release notes
PAX/ Lean 4 proof-module surfaces
src/ CUDA/PTX/Futhark source surfaces
backends/ License-policy backend
docs/ Institutional, user, architecture, and GTM documentation
demo/ Demonstration package
export_training_data.py Training-data exporter
train.py QLoRA training script
run_training.sh Training launcher
Modelfile Ollama packaging template
MODEL_CARD.md Model-card draft
DATASET_CARD.md Dataset-card draft
SOVEREIGN_NODE_KEY.md Node-key and seal policy
CONTRIBUTING.md Contribution guidance

Release Gates

The package may be published as an institutional repository release when:

  • Version files and release notes are present.
  • README states the PAX axiom basis and governance rules.
  • License text matches LICENSE.tri.
  • GitHub About metadata and topics identify the institutional scope.
  • Release notes do not overclaim artifact-specific runtime verification.
  • python export_training_data.py completes in the release environment.

Generated kernels require additional artifact-specific gates:

  • Lean/Lake proof check under the declared PAX axiom basis.
  • CUDA/PTX compiler output for the target architecture.
  • Runtime comparison against a functional reference on the target hardware.
  • License path selection and node-key/seal policy, when production use applies.

GitHub Topics

Recommended repository topics for v1.0.0:

  • pax-coder
  • formal-verification
  • lean4
  • cuda
  • ptx
  • futhark
  • gpu-kernels
  • proof-carrying-code
  • tensor-cores
  • ampere
  • deepseek-coder
  • qlora
  • sovereign-compute
  • verified-kernels
  • model-training

Release Artifact

The GitHub release should use tag v1.0.0.

Release assets are the automatic source archives generated by GitHub unless a separate model artifact, GGUF file, dataset export, or signed binary package is explicitly attached later.

v1.0.0 Packaging Evidence

Observed on Windows:

Total unique examples: 10
train: 9 examples
val: 0 examples
test: 1 examples

Generated build/ outputs are package build products and are not part of the source release unless explicitly attached as release assets.