Commit History
-
Re-measure the cleaning effect with this model instead of its sibling
Verified signature6457f29 -
Correct the cross-lingual figure and attribute the cleaning measurement
Verified signature02bda7c
-
Correct the training hyperparameters, document the input format and its two surprises
Verified signatureThe card listed learning rate 2e-5, weight decay 0.01 and batch size 32; the run used 1.5e-5, 0.02 and an effective batch of 64 through two gradient accumulation steps, with a 0.15 warm-up ratio and polynomial decay. It also said code blocks are replaced by their language name and that class 55 was filtered out for lack of samples: neither is true. Code bodies reach the model, and 55 is simply absent from the training data. Added what users most need and the card did not say: the input must go through the same cleaning the model was fine-tuned on, and that cleaning drops everything after the first install command, which cost 12.7% of raw READMEs more than half their text. The training corpus was built the same way, so the behaviour is load-bearing. The Quick Start now passes truncation=True; without it any README over 512 tokens raised an index error. Limitations now state that non-English READMEs are not read (28% self-consistency across languages), that about half of production rep…a09e369
-
Document how to avoid Hub requests on repeated inference
Verified signaturefrom_pretrained revalidates against the Hub on every call, so batch jobs split into chunks issue requests per run even when cached, inflating the download counter. Documents local_files_only / HF_HUB_OFFLINE, and clarifies that Mac GPU OOM comes from untruncated README text rather than batch size.59bc1fd -
Fix broken model id in examples, and set the device in Quick Start
Verified signatureThe Usage examples referenced alexanderquispe/naics-github-classifier, which returns HTTP 401 — copying the Quick Start failed before reaching any inference. They now point at this repo. Both examples also select cuda/mps/cpu explicitly, so a Mac user gets the GPU instead of a silent ~40x slower CPU fallback. Both snippets were executed as published.785ad04 -
Add Apple Silicon (MPS) usage section
Verified signaturePyTorch does not auto-select mps, so Mac users silently run on CPU at ~4 rows/s instead of 177. Documents the device+dtype setup, length-sorted batching, measured throughput on an M5 Max, and why fp16 is safe above the 0.8 confidence threshold.290a4fd
-
Update README.md
Verified signature4e7de42
-
Update README.md
Verified signatureecd7012 -
Update README.md
Verified signature915e4bd -
Upload label_mappings.json with huggingface_hub
Verified signaturec6ca83b -
Upload tokenizer
Verified signature884e3fb -
Upload RobertaForSequenceClassification
Verified signature3cdfac3 -
initial commit
Verified signature62e7f8b