metadata
library_name: peft
tags:
- text-generation
inference: true
widget:
- text: >
mask all personally identificable information PII including person names
for the text below: John Doe, currently lives at 1234 Elm Street,
Springfield, Anywhere 12345. He can be reached at [email protected] or at
the phone number 555-123-4567. His social security number is 123-45-6789,
and he has a bank account number 9876543210 at Springfield Bank. John
attended Springfield University where he earned a Bachelor's degree in
Computer Science. He now works at Acme Corp and his employee ID is 123456.
John's medical record number is MRN-001234, and he has a history of asthma
and high blood pressure. His primary care physician is Dr. Jane Smith,
who practices at Springfield Medical Center. His recent blood test results
show a cholesterol level of 200 mg/dL and a blood glucose level of 90
mg/dL.
base_model: meta-llama/Llama-2-7b-hf
install the required packages : peft, transformers, BitsAndBytes , accelerate
import transformers
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from torch import cuda, bfloat16
base_model_id = 'meta-llama/Llama-2-7b-hf'
device = f'cuda:{cuda.current_device()}' if cuda.is_available() else 'cpu'
bnb_config = transformers.BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type='nf4',
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=bfloat16
)
hf_auth = "hf_your-huggingface-access-token"
model_config = transformers.AutoConfig.from_pretrained(
base_model_id,
use_auth_token=hf_auth
)
model = transformers.AutoModelForCausalLM.from_pretrained(
base_model_id,
trust_remote_code=True,
config=model_config,
quantization_config=bnb_config,
device_map='auto',
use_auth_token=hf_auth
)
config = PeftConfig.from_pretrained("Ashishkr/PII-Masking")
model = PeftModel.from_pretrained(model, "Ashishkr/PII-Masking").to(device)
model.eval()
print(f"Model loaded on {device}")
tokenizer = transformers.AutoTokenizer.from_pretrained(
base_model_id,
use_auth_token=hf_auth
)
def remove_pii_info(
model: AutoModelForCausalLM,
tokenizer: AutoTokenizer,
prompt: str,
max_new_tokens: int = 128,
temperature: float = 0.92):
inputs = tokenizer(
[prompt],
return_tensors="pt",
return_token_type_ids=False).to(device)
max_new_tokens = inputs["input_ids"].shape[1]
# Check if bfloat16 is supported, otherwise use float16
dtype_to_use = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
with torch.autocast("cuda", dtype=dtype_to_use):
response = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
temperature=temperature,
return_dict_in_generate=True,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.pad_token_id,
)
decoded_output = tokenizer.decode(
response["sequences"][0],
skip_special_tokens=True,
)
return decoded_output[len(prompt) :]
prompt = """
Input: "John Doe, currently lives at 1234 Elm Street, Springfield, Anywhere 12345.
He can be reached at [email protected] or at the phone number 555-123-4567. His social security number is 123-45-6789,
and he has a bank account number 9876543210 at Springfield Bank. John attended Springfield University where he earned
a Bachelor's degree in Computer Science. He now works at Acme Corp and his employee ID is 123456. John's medical record number
is MRN-001234, and he has a history of asthma and high blood pressure. His primary care physician is Dr. Jane Smith,
who practices at Springfield Medical Center. His recent blood test results show a cholesterol level of 200 mg/dL and a
blood glucose level of 90 mg/dL.
" Output: """
# You can use the function as before
response = remove_pii_info(
model,
tokenizer,
prompt,
temperature=0.7)
print(response)