UI_Survey

The ui_survey dataset is a collection of datasets derived from unemployment insurance state survey. Specifically, it is derived from the Department of Labor's (DoL) Comparison of State Unemployment Insurance Laws which they publish annually. (This dataset is based off of the 2023 survey.) The dataset was designed for the benchmarking of both LLMs' retrieval and statutory reasoning ability.

The first dataset is ui_tables which aims to store the information from the DoL's survey. ui_tables contains 82 rows corresponding roughly to the .

Dataset Structure

>>> load_dataset('reglab/ui_survey', 'ui_tables')
Dataset({
    features: ['table_id', 'table_name', 'table_description', 'notes', 'table_data', 'chapter', 'headers', 'prompt_context', 'prompts'],
    num_rows: 82
})

>>> load_dataset('reglab/ui_survey', 'ui_qas')
Dataset({
    features: ['idx', 'table_id', 'column_num', 'jurisdiction', 'question_context', 'question', 'answer'],
    num_rows: 3400
})

Data Instance

An instance of ui_survey (here, index 13) looks like

{
  "table_id": "2-5",
  "table_name": "BENEFIT-WAGE-RATIO FORMULA STATES",
  "table_description": "Details of years used for benefits and payrolls in benefit-wage-ratio formula states",
  "notes": "",
  "table_data": "[{'state': 'DE', 'years_of_benefits_used': 'Last 3 years', 'years_of_payrolls_used': 'Last 3 years'}, {'state': 'OK', 'years_of_benefits_used': 'Last 3 years', 'years_of_payrolls_used': 'Last 3 years'}]", 'chapter': '2', 'headers': "{'state': string[python], 'years_of_benefits_used': string[python], 'years_of_payrolls_used': string[python]}", 'prompt_context': 'BENEFIT-WAGE-RATIO FORMULA—The benefit-wage-ratio formula is significantly different from the other formulas. It makes no attempt to measure all benefits paid to the workers of individual employers. The relative experience of employers is measured by the separations of workers which result in benefit payments, but the duration of their benefits is not a factor. The separations, weighted with the wages earned by the workers with each base period employer, are recorded on each employer’s experience rating record as “benefit wages.” Only one separation per beneficiary per benefit year is recorded for any one employer. The index which is used to establish the relative experience of employers is the proportion of each employer’s payroll which is paid to those workers who become unemployed and receive benefits (i.e., the ratio of an employer’s benefit wages to total taxable wages). The ratio of total benefit payments and total benefit wages, known as the state experience factor, means that, on average, the workers who drew benefits received a certain amount of benefits for each dollar of benefit wages paid and the same amount of taxes per dollar of benefit wages is needed to replenish the fund. The total amount to be raised is distributed among employers in accordance with their benefit-wage-ratios; the higher the ratio, the higher the rate. \nIndividual employer rates are determined by multiplying the employer’s experience factor by the state experience factor. The multiplication is facilitated by a table, which assigns rates that are the same as, or slightly more than, the product of the employer’s benefit-wage-ratio and the state factor.', 'prompts': "['Given the description above, what are the years of benefits used to calculate the benefit-wage ratio in {jurisdiction}?', 'Given the description above, what are the years of payrolls used to calculate the benefit-wage ratio in {jurisdiction}?']"
  "chapter": "2",
  "headers": "{'state': string[python], 'years_of_benefits_used': string[python], 'years_of_payrolls_used': string[python]}",
  "prompt_context": "BENEFIT-WAGE-RATIO FORMULA—The benefit-wage-ratio formula is significantly different from the other formulas. It makes no attempt to measure all benefits paid to the workers of individual employers. The relative experience of employers is measured by the separations of workers which result in benefit payments, but the duration of their benefits is not a factor. The separations, weighted with the wages earned by the workers with each base period employer, are recorded on each employer’s experience rating record as “benefit wages.” Only one separation per beneficiary per benefit year is recorded for any one employer. The index which is used to establish the relative experience of employers is the proportion of each employer’s payroll which is paid to those workers who become unemployed and receive benefits (i.e., the ratio of an employer’s benefit wages to total taxable wages). The ratio of total benefit payments and total benefit wages, known as the state experience factor, means that, on average, the workers who drew benefits received a certain amount of benefits for each dollar of benefit wages paid and the same amount of taxes per dollar of benefit wages is needed to replenish the fund. The total amount to be raised is distributed among employers in accordance with their benefit-wage-ratios; the higher the ratio, the higher the rate. \nIndividual employer rates are determined by multiplying the employer’s experience factor by the state experience factor. The multiplication is facilitated by a table, which assigns rates that are the same as, or slightly more than, the product of the employer’s benefit-wage-ratio and the state factor.",
  "prompts": "['Given the description above, what are the years of benefits used to calculate the benefit-wage ratio in {jurisdiction}?', 'Given the description above, what are the years of payrolls used to calculate the benefit-wage ratio in {jurisdiction}?']"
}

where in this instance

>>> pd.DataFrame.from_records(table_data[13])
  state years_of_benefits_used years_of_payrolls_used
0    DE           Last 3 years           Last 3 years
1    OK           Last 3 years           Last 3 years

An instance of ui_qas looks like

{
  "idx": 0,
  "table_id": "1-1",
  "column_num": 0,
  "jurisdiction": "Alaska",
  "question_context": "EMPLOYERS As mentioned above, one of the basic factors in determining coverage is whether services are performed for employers. The coverage provisions of most state laws use the terms “employing unit” and “employer” to make the distinctions needed to address this issue. “Employing unit” is the more generic term, it applies to any one of several specified types of legal entities that has one or more individuals performing service for it within a state. An “employer” is an employing unit that meets the specific requirements of UI law. Accordingly, services provided for an “employer” are covered, and, as a result, an employer is subject to UI tax liability and its workers accrue rights to receive UI benefits. \nFor federal UI purposes, whether an employing unit is an employer depends on the number of days or weeks a worker is employed or the amount of the employing unit’s quarterly or yearly payroll. Except for agricultural labor and domestic service, FUTA applies to employing units who paid wages of $1,500 or more during any calendar quarter in the current or immediately preceding calendar year, or to employing units with one or more workers on at least one day in each of 20 different weeks during the current or immediately preceding calendar year. About half of the states use this federal definition. \nThe following table provides information on which employing units are considered employers in each state that uses a definition other than the one in FUTA.",
  "question": "Given the description above, what is either the minimum period of time or the payroll required for an employing unit to be considered an employer in Alaska?",
  "answer": "Any time"
}

Data Fields

ui_tables

• table_id: str, ID of table written as "[chapter no.]-[table no. within chapter]" (derived from DoL report)
• table_name: str, title of the table as noted on the DoL report
• table_description: str, description of the table
• note: str, any footnotes and other key information stored for this table
• table_data: str, the data from this table as derived from the DoL report
  • eval(...) should produce a list of dictionaries corressponding to the table
• chapter: str, chapter the table appears in (TODO: replace with descriptor)
• header: str, string representation of headers with their respective datatypes
  • eval(...) should produce a dictionary
• prompt_context: str, context for the table and prompts associated with the table (intended to precede the prompts in queries)
• prompts: str, prompts to be used
  • It should be that len(prompts) == len(headers)
  • These prompts are used to derive the questions in ui_tables

ui_qas

• idx: int, index of question-answer pair
• table_id: str, ID of table the question corresponds to, written as "[chapter no.]-[table no. within chapter]" (derived from DoL report)
• column_num: str, the column number that the (relative to headers)
• jurisdiction: jurisidction (state or territory)
• question_context: str, context for the table in question (intended to precede the question in queries)
• question: str, question whose response is intended to fill in the entry corressponding to table_id and column_num for jurisidiction
• answer: str, answer to the question (from DoL report)