| license: other | |
| # syssec-utd/pylingual-shared-statement-v1 | |
| Universal (all-version) PyLingual statement backbone for the shared-backbone + per-version LoRA | |
| study, trained on Python 3.6-3.14 with 3.15 held out. `last-checkpoint/` holds the latest | |
| resumable Trainer checkpoint while training runs; the repo root holds the final model. | |
| - Data: [`syssec-utd/pylingual-shared-encoded-v1`](https://huggingface.co/datasets/syssec-utd/pylingual-shared-encoded-v1) `universal/statement` at revision `0f70ef5ada329b652a6cd67502244db733f98109` | |
| (50000 source files, 4663180 encoded rows, tokenizer.json sha256 `576881c7af5310402326eff7bf5584ad6e2f88dcfd4c998d4e8b6a26f243153f`) | |
| - Initialization: `/home/henryh/pylingual-runs/hub/Salesforce--codet5-base` | |
| - 2 epochs, global batch 72 (24 x 3 accumulation x 1 GPU), | |
| learning rate 2e-05, weight decay 0.01, FP16, seed 20261020 | |
| - Code: https://github.com/quinntyx/pylingual branch `feat-shared-backbone-lora`, commit `cb0c7b9fa5810ed5c1c09c1fe7a17c0a1eaf5606` | |
| `run/provenance.json` records every training argument and the full data manifest. | |