primitive-operation-painter-weight

Primitive Operation Painter — EMA inference weights

This local package contains EMA inference weights for a custom PyTorch autoregressive model that predicts sequences of drawing primitives. It is prepared for the future Hugging Face model repository primitive-operation-painter-weight; this directory has not been uploaded.

Architecture

• GeometrizeGPT, vocabulary size 2961
• 24 Transformer layers, hidden size 1024, 16 attention heads
• 144-step context: 10 prefix steps + 134 predicted steps
• Nine discrete tokens per drawing step; token layout geometrize_256_v1
• EMA weights only, stored in model.safetensors

Local loading

Clone or otherwise obtain the accompanying primitive-operation-painter source code, then run:

from pretrained import load_pretrained

model, config, model_dir = load_pretrained("/path/to/primitive-operation-painter-weight")

The loader uses this package's config.json. The accompanying public source repository is configured for this same 144-step model; do not load these weights with code configured for another context length.

Training metadata

• Completed epoch: 3
• Global optimizer steps: 5397
• Recorded smoothed training loss: 3.813381

The complete optimizer state and training data are intentionally excluded. The data is not redistributed with this release; users must ensure they have the necessary rights for any data they use.

License

MIT. This applies to the released code and EMA weights. Verify that your own input and training data may be used for your intended purpose.