Persona Bias Counterfactuals

This dataset contains counterfactual examples used for persona-bias circuit discovery and intervention experiments.

Repository Layout

• Hugging Face dataset config = model
• Hugging Face dataset split = counterfactual strategy
• task and axis are columns, not separate dataset configs

data/<model>/<strategy>.jsonl.gz
manifest.jsonl

Strategies

Split	Meaning
original	Full original counterfactual set derived from persona scans
axis100_dir20_balanced	Main paper counterfactual subset; capped by axis/direction budget
reverse_axis100_dir20_balanced	Reverse-direction counterfactual subset for reverse-circuit analysis

The main counterfactual strategy used in the paper is:

axis100_dir20_balanced

Row Counts by Model

Model	Rows
Llama-3.1-8B-Instruct	78694
Llama-3.2-3B-Instruct	146819
Qwen2.5-7B-Instruct	279910

Row Counts by Strategy

Strategy	Rows
axis100_dir20_balanced	11711
original	482001
reverse_axis100_dir20_balanced	11711

Load Dataset

from datasets import load_dataset

ds = load_dataset(
    "PersonaBias/counterfactuals",
    "Llama-3.2-3B-Instruct",
    split="axis100_dir20_balanced",
)

print(ds)
print(ds[0])

Filter by Task and Axis

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

print(len(arc_emotion))
print(arc_emotion[0])

Filter by direction if the column is available in the selected split:

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

Check available values:

print(ds.column_names)
print(sorted(set(ds["task"])))
print(sorted(set(ds["axis"])))

if "direction" in ds.column_names:
    print(sorted(set(ds["direction"]))[:20])

Important Schema Note

The raw counterfactual rows come from multiple tasks and model outputs. Some fields have different types across tasks, for example numeric predictions in one task and string predictions in another.

To make the Hugging Face dataset load reliably, every direct source field is stored as a string. Structured values are stored as JSON strings. The full original row is preserved in:

source_json

This means all fields used in the experiment are still available directly as columns when possible, and fully available through source_json.

Common Direct Columns

Every row includes stable metadata columns:

model
task
axis
strategy
counterfactual_strategy
source_file
summary_file
source_json

Common counterfactual fields are also directly accessible when present in the source rows:

sample_id
direction
clean_value
corrupt_value
clean_candidate
corrupt_candidate
clean_candidate_prediction
corrupt_candidate_prediction
clean_candidate_correct
corrupt_candidate_correct

The exact available columns can be inspected with:

print(ds.column_names)

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 = ds[0]

source_row = json.loads(row["source_json"])

sample_id = as_int(row["sample_id"]) if "sample_id" in row else None
clean_correct = (
    as_bool(row["clean_candidate_correct"])
    if "clean_candidate_correct" in row
    else None
)

print(source_row.keys())

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_counterfactual_row(row):
    parsed = dict(row)
    parsed["source"] = json.loads(row["source_json"])

    for key in [
        "sample_id",
        "clean_candidate_prediction",
        "corrupt_candidate_prediction",
    ]:
        if key in row:
            parsed[f"{key}_int"] = as_int(row[key])

    for key in [
        "clean_candidate_correct",
        "corrupt_candidate_correct",
    ]:
        if key in row:
            parsed[f"{key}_bool"] = as_bool(row[key])

    return parsed

Example Analysis

from collections import Counter
from datasets import load_dataset

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

ds = load_dataset(repo, model, split="axis100_dir20_balanced")

print("Rows:", len(ds))
print("Task counts:", Counter(ds["task"]))
print("Axis counts:", Counter(ds["axis"]))

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)
        print(task, axis, len(subset))

Available Tasks and Axes

Main tasks:

arc_easy
arc_challenge
ethics
safety
sst2

Axes:

emotion
gender
race
religion

Notes:

• gsm8k is not part of the main paper setting.
• axis100_dir20_balanced is the recommended split for main circuit discovery.
• reverse_axis100_dir20_balanced is intended for reverse-circuit diagnostics.
• original is larger and includes the full derived counterfactual pool.

Manifest

manifest.jsonl records the source summary/data file and row count for each model/task/axis/strategy group used to build the upload.