223344 / README.md
zaaabik's picture
README: one-command download into cct-reproduce/data
64e2dac verified
|
Raw
History Blame Contribute Delete
2.56 kB
metadata
license: cc-by-4.0
pretty_name: CCT  Class-Centroid Trajectories, pre-computed features
tags:
  - uncertainty-estimation
  - misclassification-detection
  - transformers
  - text-classification

CCT — pre-computed features, predictions, results, and encoders

Companion data for Class-Centroid Trajectories for Admissible Misclassification Detection in Fine-Tuned Transformer Classifiers. The repository is laid out exactly as the code expects, so it drops into the cct-reproduce package with a single command and no moving of files:

hf download zaaabik/223344 --repo-type dataset --local-dir cct-reproduce/data
# older CLI: huggingface-cli download zaaabik/223344 --repo-type dataset --local-dir cct-reproduce/data
pytest cct-reproduce/tests/          # 46 tests, reproduces the paper numbers

Download only what you need (each is self-contained):

# CCT-25 + predictions only (~2 GB) — enough for every CCT table/figure
hf download zaaabik/223344 --repo-type dataset --local-dir cct-reproduce/data --include "features/*" "results/*"
# one configuration only
hf download zaaabik/223344 --repo-type dataset --local-dir cct-reproduce/data --include "features/roberta_cola/*"

Layout

Path Size Contents
features/{cfg}/ ~2 GB cct_{split}.npz (CCT-25), cls_{begin,mid,last}_{split}.npz, predictions_{split}.parquet
sv1/{cfg}/ ~7 GB AttnTopo per-family features: {ripser,template,graph,toktopo_pd,toktopo_graph,intra_attn,punct,cls_last,cls_mid,cls_begin}_{split}.npz (14,608 columns in total)
results/ small cached per-paper-table CSVs
checkpoints/{cfg}/ ~2.5 GB fine-tuned encoders (seed 0), for re-extracting features from scratch

{cfg} is one of {roberta,electra}_{cola,sst5,toxigen,newsgroups,goemotions} plus roberta_yelp and electra_yelp; {split} is train, validation, or test. Splits are the ones the encoders were fine-tuned with, so there is no train→validation leakage in detector training.

Note when re-extracting features from checkpoints/

The weights are stored in bfloat16. Load them in float32 — with transformers>=5, from_pretrained follows the checkpoint dtype by default and bf16 inference shifts logits by up to ~0.08, which flips a few predictions and perturbs the depth-based features:

model = AutoModelForSequenceClassification.from_pretrained(path, dtype=torch.float32)

With float32 inference the released encoders reproduce the shipped predictions_*.parquet logits to ~1e-5.