JordanTensor's picture
Update README.md
ddb2901 verified
metadata
dataset_info:
  features:
    - name: hard_text
      dtype: string
    - name: profession
      dtype: int64
    - name: gender
      dtype: int64
    - name: ai_correct
      dtype: bool
  splits:
    - name: train
      num_bytes: 1612718.7804411764
      num_examples: 4453
    - name: test
      num_bytes: 524052.11661764706
      num_examples: 1447
  download_size: 1251600
  dataset_size: 2136770.8970588236
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

Dataset Card for "bias_in_bios_verified_software_devs_only"

A dataset of software engineer bios, filtered from the LabHC/bias_in_bios dataset, by selecting the software engieer class and using gpt-4o-mini to filter out bios which don't look like software engineers.

For example 900 / 6800 rows like this have been removed:

He has been playing guitar and devoting himself to music for many years. There is no question that music is his life. He has performed in music houses and clubs with musicians from Taiwan and other countries. His passion and desire to learn more about music and guitar influenced him to enroll at the Musicians Institute in LA next year. He is an enthusiast of blues, swing jazz and rockabilly.

His first year competing in American Ninja Warrior was in Season 6. He completed the course in the Dallas Qualifier. Jo Jo seemed to take his time through each obstacle. During the Ring Toss he missed a couple of pegs but was able to recover in order to finish. Jo Jo was able to make it up the Warped Wall in his first attempt finishing in eighteenth place.

The prompt for this filtering was

prompt = f"""I'm going to give you a biography, and I want you determine if it belongs to a software engineer. 

Biography:
"{bio}"

Is this person likely to be a software engineer? Please respond only with "Yes." if you believe they are likely to be a software engineer, or "No." otherwise.
"""

output = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {
            "role": "user",
                "content": prompt
        }
    ],
)

Warning: the dataset has also been shuffled, and modified to have a train test split of 4453 / 1447.