GeoRefine TBE tools

georefine-verify independently compares a GeoRefine TBE bundle with the published parent model, byte for byte. georefine-export restores a TBE bundle to ordinary Hugging Face safetensors without downloading or reading the parent. Both tools use NumPy and the Python standard library, and run on CPU.

pip install .
georefine-export --bundle Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-TBE \
  --output ./Qwen3.8-27B-GeoRefine-BF16 \
  --expect-manifest-sha256 10028416e07c802b47e4dc93a86e9f0dd35ce6021f0e07b96c70bd9cb7a72bc3

The output has model.safetensors.index.json, standard safetensors shards, and the original config, tokenizer and vision files. Load it with a current Transformers release compatible with the parent model:

from transformers import AutoModelForMultimodalLM, AutoProcessor
model = AutoModelForMultimodalLM.from_pretrained("./Qwen3.8-27B-GeoRefine-BF16")
processor = AutoProcessor.from_pretrained("./Qwen3.8-27B-GeoRefine-BF16")

The export is BF16: plan for about 56 GB of output storage, plus one codec shard of scratch space when downloading from the Hub, and enough memory or GPU capacity to run Qwen3.8-27B. This path establishes portable model compatibility; it does not retain the compressed model's runtime memory or speed advantages. The native TBE engine is a separate optimized backend. A full 27B export and Transformers load remain release gates, beyond the existing small-fixture tests.