Accuknoxtechnologies/PII-Qwen3.5-2B-v8

Qwen PII Guard (merged)

Fine-tuned from Qwen/Qwen3.5-2B to detect personally-identifiable information in user prompts and emit a single JSON object listing the values found in each of 15 categories.

Output schema:

{"is_valid": true,
 "category": {"Name": ["John Doe"], "Email": ["john@example.com"]}}

is_valid is false and category is {} when the prompt contains no PII.

Categories

name, email, phonenumber, address, date, nationalid, passportnumber, driverslicense, taxid, cardnumber, bankaccount, credentials, ipaddress, username

Evaluation (transformers reference path)

• —test rows: 200 (held-out, from test_dataset_pii.csv)
• —is_valid accuracy: 1.0000
• —category key-set accuracy: 0.9350
• —category value-set accuracy: 0.8300
• —binary F1 (is_valid): 1.0000 (P=1.000 R=1.000)
• —macro F1 over categories (key-presence): 0.9791
• —macro F1 over categories (value-set): 0.9529
• —parse errors: 0/200

Binary confusion matrix (positive = "contains PII"):

Per-category KEY-presence (did the model emit this category at all?):

Per-category VALUE-set (did the exact strings match within the category?):

Latency (transformers, single-prompt, greedy decoding):

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Accuknoxtechnologies/PII-Qwen3.5-2B-v8")
model = AutoModelForCausalLM.from_pretrained("Accuknoxtechnologies/PII-Qwen3.5-2B-v8", torch_dtype="auto", device_map="auto")

prompt = "Please contact me at jane@example.com or +1 415 555 0100."
msgs = [
    {"role": "system", "content": "<see SYSTEM_MSG in train_qwen_pii.py>"},
    {"role": "user", "content": prompt},
]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device),
                    max_new_tokens=512, do_sample=False, pad_token_id=tok.pad_token_id)
print(tok.decode(out[0], skip_special_tokens=True))

Evaluation — vLLM serving (merged model, text-only)

Same 200 held-out prompts, served through vLLM `0.21.0` instead of the transformers .generate() loop. Greedy decoding, dtype bf16, enable_prefix_caching=True, enable_chunked_prefill=True. This reflects production serving accuracy + latency.

• —JSON parse errors: 0/200 (0.0%)

Accuracy (vLLM)

Confusion matrix — binary is_valid (vLLM)

Per-category key-presence (vLLM)

vLLM inference latency (single-stream, batch = 1)

vLLM throughput (single batched submit)

• —Prompts/sec: 27.73
• —Output tokens/sec: 1569.0
• —Input tokens/sec: 35596.5
• —Batched wall time for all 200 prompts: 7.21 s

Card generated at 2026-05-31 07:39 UTC. Adapter weights: `Accuknoxtechnologies/PII-Qwen3.5-2B-v8`.