README.md

metadata

license: other
pretty_name: PyLingual shared-backbone encoded training data

syssec-utd/pylingual-shared-encoded-v2

Tokenized Arrow datasets for the shared-backbone + per-version LoRA PyLingual study, exactly as the trainers load them (datasets.load_from_disk("<path>/dataset"); <path>/manifest.json records the journal, tokenizer and registry hashes the trainers verify). Statement inputs and targets are truncated at 256 tokens; segmentation uses 512-token windows with stride 128.

Built from:

• syssec-utd/pylingual-shared-tokenizer-v1 at revision d23233a4bb237dcc3b1660e46a42f9189eaf0f6c (tokenizer)
• syssec-utd/pylingual-shared-backbone-corpus-v2 at revision 3dc69d6e21e5ffa3ce461caf7d7bb791bb2f603a (backbone-corpus)
• syssec-utd/pylingual-shared-adapter-corpus-v2 at revision 95f11e98728b0cf350a684e202af02e5398781f0 (adapter-corpus)
• Code: https://github.com/quinntyx/pylingual branch feat-shared-backbone-lora, commit 87226a491a9ee76670601a6b3e8a37e8187241c5

Path	Task	Scope	Split	Source files	Encoded rows	Opcode bindings
adapter/3.14/statement/train	statement	adapter	train	100000	9378298	{}
adapter/3.14/statement/validation	statement	adapter	validation	1000	95474	{}
universal/statement	statement	universal	train	100000	9289674	{}