GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Liquor-06
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Liquor-06
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
The model answers with a JSON list of {"type", "value"} entries, each value copied verbatim from the
text. Ask exactly as it was trained: the system prompt below, the text after Text:\n, thinking off,
greedy decoding.
python
import jsonfrom transformers import AutoModelForCausalLM, AutoTokenizer name = "Liquor-06/Qwen3-0.6B-PII-distillops"tok = AutoTokenizer.from_pretrained(name)model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto") SYSTEM = ( 'List every piece of personal information (PII) in the text so that it can be masked.\n' '- Copy each value exactly as it appears in the text, character for character.\n' '- Words of the same category that stand next to each other form one entry: a full name with its title ("Dr Ana Maria Lee"), a street with its number ("12 Oak Street"), a city with its postal code ("Boston 02108"). Pieces separated by other words or by punctuation (a comma, a hyphen) are separate entries.\n' '- Every city or town name counts, even on its own. For an age, list only the number ("27", not "27 years old"); gender and sex words count too.\n' '- List a value once even if it appears several times.\n' 'Answer with a JSON list only: [{"type": "<CATEGORY>", "value": "<exact text>"}, ...], or [] if the text has no personal information.\n' '\n' 'Categories:\n' "- NAME: a person's name, with any title (Mr, Dr, ...)\n" '- CONTACT: an email address or phone number\n' '- ADDRESS: a street name, building number, city or postal code\n' "- ID_NUMBER: an ID card, passport, driver's license, social security or tax number\n" '- CARD_NUMBER: a payment card number\n' '- DATE: a date\n' "- DEMOGRAPHIC: a person's age, gender or sex") text = "Dear Mr John Smith, your appointment on 12 March 2024 is confirmed. Call us on 020 7946 0958."messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": f"Text:\n{text}"}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=False)entities = json.loads(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))print(entities) # e.g. [{"type": "NAME", "value": "Mr John Smith"}, {"type": "DATE", ...}, ...] masked = textfor e in sorted(entities, key=lambda e: -len(e["value"])): # longest first masked = masked.replace(e["value"], f"[{e['type']}]")print(masked)Pass do_sample=False: generation_config.json keeps Qwen3's sampling defaults, and all scores above
are greedy. Generation stops at <|im_end|>. With vLLM: vllm serve Liquor-06/Qwen3-0.6B-PII-distillops,
then send the same messages with chat_template_kwargs: {"enable_thinking": false} and temperature: 0.
Categories: NAME (with any title), CONTACT (email, phone), ADDRESS (street, building number, city,
postal code), ID_NUMBER (ID card, passport, driver's license, social security, tax number),
CARD_NUMBER, DATE, DEMOGRAPHIC (age, gender, sex). Adjacent words of one category form one entry
("Dr Ana Maria Lee", "12 Oak Street").
deepseek-flash, 3 worked examples) labelled the
10,000 texts. An answer was accepted by rules alone (known categories, every value copied verbatim,
plausible formats, no 9+ digit number left unmasked), as for a company's own unlabelled text: 9,681
texts (96.8%) passed. Gold labels of 100 texts were used only to write the teacher's instructions.Commands, configs and the full write-up: github.com/Zihao-Wu06/distillops.
Apache-2.0, as for the base model. OpenPII-1M is CC-BY-4.0 (ai4privacy); the training labels came from the DeepSeek API (check DeepSeek's terms of use for your use case).
import jsonfrom transformers import AutoModelForCausalLM, AutoTokenizer name = "Liquor-06/Qwen3-0.6B-PII-distillops"tok = AutoTokenizer.from_pretrained(name)model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto") SYSTEM = ( 'List every piece of personal information (PII) in the text so that it can be masked.\n' '- Copy each value exactly as it appears in the text, character for character.\n' '- Words of the same category that stand next to each other form one entry: a full name with its title ("Dr Ana Maria Lee"), a street with its number ("12 Oak Street"), a city with its postal code ("Boston 02108"). Pieces separated by other words or by punctuation (a comma, a hyphen) are separate entries.\n' '- Every city or town name counts, even on its own. For an age, list only the number ("27", not "27 years old"); gender and sex words count too.\n' '- List a value once even if it appears several times.\n' 'Answer with a JSON list only: [{"type": "<CATEGORY>", "value": "<exact text>"}, ...], or [] if the text has no personal information.\n' '\n' 'Categories:\n' "- NAME: a person's name, with any title (Mr, Dr, ...)\n" '- CONTACT: an email address or phone number\n' '- ADDRESS: a street name, building number, city or postal code\n' "- ID_NUMBER: an ID card, passport, driver's license, social security or tax number\n" '- CARD_NUMBER: a payment card number\n' '- DATE: a date\n' "- DEMOGRAPHIC: a person's age, gender or sex") text = "Dear Mr John Smith, your appointment on 12 March 2024 is confirmed. Call us on 020 7946 0958."messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": f"Text:\n{text}"}]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=False)entities = json.loads(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))print(entities) # e.g. [{"type": "NAME", "value": "Mr John Smith"}, {"type": "DATE", ...}, ...] masked = textfor e in sorted(entities, key=lambda e: -len(e["value"])): # longest first masked = masked.replace(e["value"], f"[{e['type']}]")print(masked)