Datasets:
Patent Boundary Notice
Scope
The retokenization pipeline (retokenize_scylla.py) included in this repository converts
SentencePiece-tokenized FineWeb shards to alternative tokenizer vocabularies. The algorithm
is straightforward: decode tokens to bytes, re-encode with the target tokenizer, validate
byte-level roundtrip fidelity.
What is NOT covered
The pre-tokenized data files (.bin shards) in this dataset are standard binary token
sequences. They do not embody any patentable method — they are the output of running a
tokenizer on public data.
What MAY be covered
Techniques described in patent applications by Light Speed Up LLC related to:
- Sensitivity-guided mixed-precision quantization
- Adaptive tokenizer selection for language model compression
These techniques are implemented in training code (not in this dataset or retokenization tool). The Apache 2.0 license on the code grants patent rights for the code as-shipped.
Contact
For patent-related inquiries: mato@lightspeedup.com