PersonaBias/bias_evaluation_sets

Persona Bias Evaluation Sets

This dataset contains evaluation sets derived from full-model persona behavior. Each row is an original task sample grouped by whether changing the persona makes the model behavior biased, unbiased, or all-wrong.

Repository Layout

• Hugging Face dataset config = model
• Hugging Face dataset split = validation or test
• Behavioral subset = eval_set column

data/<model>/validation.jsonl.gz
data/<model>/test.jsonl.gz
manifest.jsonl

Evaluation Sets

eval_set	Meaning
biased	At least one persona value produces the correct answer and at least one persona value produces a wrong answer
unbiased	All persona values preserve the correct answer
all_wrong	All persona values produce wrong answers; this is saved as a separate diagnostic set

Row Counts

Model	Rows
Llama-3.1-8B-Instruct	68308
Llama-3.2-1B-Instruct	73584
Llama-3.2-3B-Instruct	73584
Qwen2.5-7B-Instruct	65316

Split	Rows
test	156280
validation	124512

Eval set	Rows
all_wrong	38031
biased	19154
unbiased	223607

Load Dataset

from datasets import load_dataset

ds = load_dataset(
    "PersonaBias/bias_evaluation_sets",
    "Llama-3.2-3B-Instruct",
    split="test",
)

print(ds)
print(ds[0])

Filter by Task, Axis, and Eval Set

task_axis_ds = ds.filter(
    lambda row: row["task"] == "arc_challenge"
    and row["axis"] == "emotion"
)

biased = task_axis_ds.filter(lambda row: row["eval_set"] == "biased")
unbiased = task_axis_ds.filter(lambda row: row["eval_set"] == "unbiased")
all_wrong = task_axis_ds.filter(lambda row: row["eval_set"] == "all_wrong")

print(len(biased), len(unbiased), len(all_wrong))

Shorter version:

biased_arc_emotion = ds.filter(
    lambda row: row["task"] == "arc_challenge"
    and row["axis"] == "emotion"
    and row["eval_set"] == "biased"
)

Important Schema Note

The source rows come from multiple tasks and persona axes. Some original fields have different types across tasks, for example numeric labels in one task and string labels in another. Some dictionary fields also have different keys across axes.

To make the Hugging Face dataset load reliably, source payload fields are stored as strings or JSON strings. The main filtering columns are ready to use directly. Some value columns are stored as strings for schema stability:

model
split
source_split
task
axis
eval_set
sample_id
task_label
groundtruth
sample

Structured fields are stored as JSON strings:

persona_results_json
predictions_by_persona_json
correctness_by_persona_json
unique_predictions_json
changed_persona_pairs_json
source_json

Parse Typed Fields

import json

def as_bool(x):
    return str(x).lower() == "true"

def as_int(x, default=None):
    text = str(x).strip()
    return int(text) if text else default

row = biased[0]

persona_results = json.loads(row["persona_results_json"])
predictions_by_persona = json.loads(row["predictions_by_persona_json"])
correctness_by_persona = json.loads(row["correctness_by_persona_json"])
source_row = json.loads(row["source_json"])

unique_prediction_count = as_int(row["unique_prediction_count"])
all_personas_wrong = as_bool(row["all_personas_wrong"])

Ready-to-Use Helper

import json

def as_bool(x):
    return str(x).lower() == "true"

def as_int(x, default=None):
    text = str(x).strip()
    return int(text) if text else default

def parse_eval_row(row):
    return {
        **row,
        "sample_id_int": as_int(row["sample_id"]),
        "unique_prediction_count_int": as_int(row["unique_prediction_count"]),
        "changed_persona_pair_count_int": as_int(row["changed_persona_pair_count"]),
        "all_personas_correct_bool": as_bool(row["all_personas_correct"]),
        "all_personas_wrong_bool": as_bool(row["all_personas_wrong"]),
        "mixed_correctness_bool": as_bool(row["mixed_correctness_across_personas"]),
        "prediction_changed_bool": as_bool(row["prediction_changed_across_personas"]),
        "persona_results": json.loads(row["persona_results_json"]),
        "predictions_by_persona": json.loads(row["predictions_by_persona_json"]),
        "correctness_by_persona": json.loads(row["correctness_by_persona_json"]),
        "unique_predictions": json.loads(row["unique_predictions_json"]),
        "changed_persona_pairs": json.loads(row["changed_persona_pairs_json"]),
        "source": json.loads(row["source_json"]),
    }

Example Analysis

from datasets import load_dataset

repo = "PersonaBias/bias_evaluation_sets"
model = "Llama-3.2-3B-Instruct"

ds = load_dataset(repo, model, split="test")

for task in sorted(set(ds["task"])):
    for axis in sorted(set(ds["axis"])):
        subset = ds.filter(lambda row: row["task"] == task and row["axis"] == axis)
        counts = {
            name: len(subset.filter(lambda row: row["eval_set"] == name))
            for name in ["biased", "unbiased", "all_wrong"]
        }
        print(task, axis, counts)

Available Tasks and Axes

Main tasks:

arc_easy
arc_challenge
ethics
safety
sst2

Axes:

emotion
gender
race
religion

Notes:

• sst2 has validation only because SST-2 test labels are unavailable.
• gsm8k scans are not included in this release.
• all_wrong is diagnostic and should not be merged into biased or unbiased.

Manifest

manifest.jsonl records each source file used to build the dataset, including model, task, split, axis, eval set, row count, and whether an expected source file was missing.