Parameter	Value
Base model	`facebook/bart-base`
Checkpoint epoch	8
Global step	12479
Saved at	2026-04-06T01:42:04.473556
Train loss	0.3834
Eval loss	0.4518
Perplexity	1.57
Token accuracy	0.4206
ERR exact	0.2925
ERR partial	0.0282
Eval samples	1918

Evaluation Metrics

Attack B — Inverter Model (best checkpoint)

Metric	Value
ERR exact	0.2925 (29.2% entities recovered verbatim)
ERR partial	0.0282
Token accuracy	0.4206
Exact sentence match	0.1528
Corpus BLEU	0.6466

ERR by adversarial strategy

Strategy	Samples	ERR exact
S2_combinatorial_pii	424	0.1958
S3_paraphrase_consistency	261	0.4521
S4_rarity_spectrum	311	0.2990
S1_entity_consistency	556	0.2878
S6_edge_cases	186	0.3548
S5_cross_entity_correlation	180	0.2278

ERR by name rarity

Rarity	Samples	ERR exact
unknown	1051	0.2931
common	128	0.3750
rare	249	0.2088
very_rare	134	0.3134
medium	356	0.3118

Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

tokenizer = AutoTokenizer.from_pretrained("JALAPENO11/pii-inverter-bart-base")
model     = AutoModelForSeq2SeqLM.from_pretrained("JALAPENO11/pii-inverter-bart-base")
model.eval()

anonymized = "Dear Ms. Anya Sharma, please update your account details."

inputs = tokenizer(anonymized, return_tensors="pt",
                   max_length=128, truncation=True)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=128, num_beams=4)
print(tokenizer.decode(out[0], skip_special_tokens=True))
# → "Dear Amyna Tharathip, please update your account details."

Disclaimer

This model is published for research purposes only to study the robustness of PII anonymisation systems. It should not be used to de-anonymise real personal data.

See attack_report.txt in the repo Files tab for the full evaluation report.