emails: list of strings
orgs: list of {"name": "...", "type": "..."}
persons: list of {"name": "...", "role": "..."}
addresses: list of {"street": "...", "city": "...", "state": "...", "zip_code": "...", "country": "...", "full_text": "..."}
phones: list of strings
bank_accounts: list of {"account_number": "...", "bank_name": "..."}
dates: list of strings

Example

Input:

SERVICE AGREEMENT
This agreement is made on June 1, 2025 between Acme Robotics Inc.,
located at 500 Market Street, Suite 12, Austin, TX 78701, and
Jane Doe, acting as an independent contractor. Contact: jane.doe@acme-robotics.com.

Output:

json

{
  "emails": ["jane.doe@acme-robotics.com"],
  "orgs": [{"name": "Acme Robotics Inc.", "type": "Inc."}],
  "persons": [{"name": "Jane Doe", "role": "independent contractor"}],
  "addresses": [{"street": null, "city": "Austin", "state": "TX", "zip_code": "78701", "country": null, "full_text": "500 Market Street, Suite 12, Austin, TX 78701"}],
  "phones": [],
  "bank_accounts": [],
  "dates": ["June 1, 2025"]
}

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "khaled12331/qwen3-4b-cuad-entities",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "khaled12331/qwen3-4b-cuad-entities",
    trust_remote_code=True
)

Full Inference Example

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

MODEL_ID = "khaled12331/qwen3-4b-cuad-entities"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

SYSTEM_PROMPT = (
    "You are a legal document entity extraction specialist.
"
    "Extract all entities from the contract into strict JSON with these keys:
"
    "emails, orgs, persons, addresses, phones, bank_accounts, dates.

"
    "CRITICAL FORMAT RULES:
"
    '- emails: list of plain strings
'
    '- orgs: list of objects like {"name": "...", "type": "..."}
'
    '- persons: list of objects like {"name": "...", "role": "..."}
'
    '- addresses: list of objects like {"street": "...", "city": "...", "state": "...", "zip_code": "...", "country": "...", "full_text": "..."}
'
    '- phones: list of plain strings
'
    '- bank_accounts: list of objects like {"account_number": "...", "bank_name": "..."}
'
    '- dates: Extract ALL dates

'
    "Output ONLY valid JSON. No markdown, no explanations."
)

contract = "YOUR CONTRACT TEXT HERE"

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": contract},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)

result = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)

Training Details

Method: LoRA (r=16, alpha=32)
Epochs: 4
Dataset: ~600-700 legal documents
Framework: TRL SFTTrainer
Base Model: Qwen/Qwen3-4B

khaled12331/qwen3-4b-cuad-entities-v2

Qwen3-4B Legal Entity Extraction

Base Model

Task

Example

Usage

Full Inference Example

Training Details

khaled12331/
qwen3-4b-cuad-entities-v2