docs/HUGGINGFACE_PUBLISHING.md

Hugging Face Publishing

Recommended Split

Option 1: Code Repo Only

Publish this repository as the harness source and keep weights elsewhere.

Best for:

• auditability
• open-source orchestration
• reproducible local verification loops

Option 2: Separate Model Repo

Store GGUF or safetensors weights in a dedicated Hugging Face model repository.

Best for:

• large artifacts
• cleaner LFS separation
• versioned checkpoints

Option 3: Space Repo

Use a Hugging Face Space for interactive theorem verification demos.

Best for:

• operator-facing UI
• public demos
• benchmark explorer

Pre-Publish Checklist

• replace placeholder model IDs
• pin Docker image tags and Lean toolchain versions
• confirm license for model weights
• add real benchmark metadata
• scrub local absolute paths
• add reproducible example commands
• decide whether to publish Granite, Llemma, or both as supported backends

Files Already Added For HF Readiness

• .gitattributes
• hf/README.md
• code/infra split suitable for model companion repos