Self-contained PyTorch example, escaped regex ranges

English comments: the published copy is read internationally

Table shows both formats and their real sizes

File table and a snippet per format

Document the PyTorch path

PyTorch weights as safetensors, with the architecture module

Add the pip line; Colab needs it

Apache-2.0, matching the teacher models

Card: state corpus provenance and what is released

English card: live demo up top, tested Python example

Card: full precision, no quantisation

Full precision for every rung: no quantisation anywhere

Per-channel int8 for the two large rungs, fp32 for the rest

Delete ruq-30m-int8.onnx with huggingface_hub

Delete ruq-15m-int8.onnx with huggingface_hub

Delete ruq-5m-int8.onnx with huggingface_hub

Delete ruq-2m-int8.onnx with huggingface_hub

Delete ruq-0.7m-int8.onnx with huggingface_hub

Add tokenizer_config.json and config.json

ONNX int8 ladder + tokenizer

initial commit