latent-calculator weights
ONNX graphs for the in-browser latent-calculator demo (fp32 and fp16 variants, shared embedding table, shards under 90 MB). Built by latent-calculator/web/export_onnx.py in oaustegard/experiments. Live demo: austegard.com/ai-tools/latent-calculator/.
ckpt/
The trained parts from latent-calculator/ (query heads, result encoders for every arm and epoch, both models; probe_*.json carries the chosen layer k), each under 4 MB. With web/export_onnx.py these regenerate the ONNX graphs above; with eval.py and demo.py they reproduce every number in RESULTS.md.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for austegard/latent-calculator-web
Base model
HuggingFaceTB/SmolLM2-135M