persona-data

Dataset loaders and prompt utilities for the implicit-personalization research effort, built around SynthPersona — an open synthetic-persona dataset for studying, steering, and personalizing language models.

The SynthPersona dataset

implicit-personalization/synth-persona is a fully open synthetic persona dataset (~1.41 GB, English) for research on implicit personalization, persona steering, and persona-grounded evaluation.

• 1,000 personas built from structured seed attributes and expanded into biographies, interview transcripts, and supporting statements, plus a baseline_assistant control.
• 788k QA rows across three axes:
  • type: explicit (supported by a seed/interview/statement) vs. implicit (inferred from the biography).
  • scope: individual (one persona) vs. shared (same item across all personas, directly comparable).
  • item_type: FRQ (free-response, for training) vs. MCQ (multiple-choice, for evaluation).
• Shared MCQ banks: 418 implicit + 57 explicit items reused across personas, with a curated study_model_evaluable_v1 subset (231 items) for 7B-scale evaluation.
• 18 topic groups (e.g. future_hopes_and_values, stress_coping_and_support) for sliced analyses.
• Leakage-aware splits: each MCQ tracks its source FRQs/seeds (bank_id, related_frq_qids), so FRQ-train / MCQ-test splits avoid contamination.

See the dataset card for the full schema.

Installation

pip install persona-data    # or: uv add persona-data

The dataset is downloaded from Hugging Face on first use and cached locally.

Quick start

from persona_data.synth_persona import SynthPersonaDataset
from persona_data.prompts import format_prompt, format_messages

dataset = SynthPersonaDataset()
persona = dataset[0]

system_prompt = format_prompt(persona, "biography")
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Where did you grow up?"},
]

# Leakage-aware split: individual FRQs for train, shared MCQs for test.
train_qa, test_qa = dataset.train_test_split(persona.id)

# Slice by topic or by curated evaluation subset.
religion = dataset.get_qa(persona.id, type="implicit",
                          topic_group_id="religion_spirituality_and_meaning")
eval_mc  = dataset.get_qa(persona.id, item_type="mcq",
                          question_set="study_model_evaluable_v1")

# Minimal-pair counterfactual: same persona, one attribute swapped in the
# templated view (binary attributes default to the opposite value).
from persona_data.templated import swap_attribute
base, swapped = swap_attribute(dataset, persona.id, "speak_other_language")

Pass sample_size=N to load only the first N personas.

What else is in the package

• SynthPersonaDataset — personas + QA pairs (docs)
• NemotronPersonasFranceDataset / NemotronPersonasUSADataset — NVIDIA persona-only datasets (docs)
• prompts — roleplay and multiple-choice formatting helpers (docs)
• templated — single-attribute counterfactual swaps on the templated view (docs)
• environment — set_seed, get_device, get_artifacts_dir

Full API reference: https://implicit-personalization.github.io/persona-data/.

Used by

• persona-vectors — activation extraction and steering
• persona-2-lora — LoRA-based persona internalization

Citation

If you use SynthPersona, please cite the dataset card and link back to this repo.