Draup/salary-normalizer

Salary Normalizer

A fine-tuned Gemma 3 270M model that parses and standardizes free-form salary text into structured JSON. Given an arbitrary salary string and a country name, it extracts currency symbol, ISO code, numeric range, and pay cadence in a single inference pass.

Model Details

Property	Value
Base model	google/gemma-3-270m
Fine-tune type	Supervised (instruction-tuned)
License	Apache 2.0
Task	Structured information extraction

Intended Use

• Extract and normalize salary mentions from job descriptions or candidate profiles.
• Standardize heterogeneous salary formats (e.g., 12 LPA, $60k–$80k/yr, €45,000 p.a.) into a consistent schema for downstream analytics or storage.

Out-of-scope use: This model is not designed for general text generation or tasks unrelated to salary parsing.

Input Format

Prompts must follow this exact template:

<start_of_turn>user
summarize salary: <SALARY_TEXT>
country: <COUNTRY_NAME><end_of_turn>
<start_of_turn>model

Examples of valid salary text:

• $60k - $80k per year
• INR 12 LPA
• €45,000 annually
• 12 to 12.5 US $ per hr

Country names must match one of the supported countries listed below.

Output Schema

The model returns a JSON object with the following fields:

{
  "currency": "$",
  "iso_code": "USD",
  "min_amount": 60000,
  "max_amount": 80000,
  "pay_rate": "ANNUALLY"
}

Field	Type	Description
currency	string	Raw currency symbol or string as it appears in the input
iso_code	string	Standardized ISO 4217 currency code
min_amount	int / float	Lower bound of the salary range (annualized or as stated)
max_amount	int / float	Upper bound of the salary range (annualized or as stated)
pay_rate	string	One of: HOURLY, DAILY, WEEKLY, BI-WEEKLY, MONTHLY, ANNUALLY, OTHERS

Note: min_amount and max_amount reflect normalized numeric values, not raw token extractions. For single-value salaries, both fields will hold the same value.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Draup/salary-normalizer"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

salary_text = "12 to 12.5 US $ per hr"
country = "United States"

prompt = (
    f"<start_of_turn>user\n"
    f"summarize salary: {salary_text}\n"
    f"country: {country}<end_of_turn>\n"
    f"<start_of_turn>model\n"
)

inputs = tokenizer(prompt, return_tensors="pt", truncation=True)

outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    do_sample=False,
    eos_token_id=tokenizer.convert_tokens_to_ids("<end_of_turn>")
)

result = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True
)
print(result)
# {"currency": "US $", "iso_code": "USD", "min_amount": 12, "max_amount": 12.5, "pay_rate": "HOURLY"}

Supported Countries

The model supports salary parsing for the following 49 countries:

Argentina	Australia	Austria	Belgium
Brazil	Canada	Chile	China
Colombia	Czechia	Denmark	Egypt
Finland	France	Germany	Hong Kong
Hungary	India	Indonesia	Ireland
Israel	Italy	Japan	Malaysia
Mexico	Netherlands	New Zealand	Norway
Pakistan	Peru	Philippines	Poland
Portugal	Romania	Russia	Saudi Arabia
Singapore	South Africa	South Korea	Spain
Sweden	Switzerland	Taiwan	Thailand
Turkey	United Arab Emirates	United Kingdom	United States
Vietnam

Limitations

• Performance may degrade on salary formats not well-represented in training data.
• Country context is used for currency disambiguation; incorrect country input may produce inaccurate iso_code or currency values.
• The model is not multilingual — salary text is expected to be in English or use standard numeric/symbol conventions.