brunomcalura's picture
Add DREAM-inspired enhanced profiles for all 200 personas
8f91598 verified
|
Raw History Blame Contribute Delete
2.39 kB
metadata
pretty_name: Persona Modality Mapping
language:
  - en
tags:
  - synthetic
  - persona
  - parquet
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/default/train-*.parquet
  - config_name: pilot_40
    data_files:
      - split: train
        path: data/pilot_40/train-*.parquet

Persona-Modality-Mapping

Persona IDs, original personas, descriptions and enhanced prompts. The default configuration contains all 200 personas and the three rewrites. The pilot configuration contains the separate 40-persona evaluation data and its enhanced description. Pilot and full-run descriptions can differ for the same ID. The reference image paths are identifiers from the source dataset.

Configurations

Configuration Train rows
default 200
pilot_40 40

Provenance

The full run used claws-lab/persona-modality at revision 1cce2b85133957584489c4367ab7d3821de98d88. The generation code revision was e7b9f2acac01db9d0efa8c09d2616d2ff7902a4c. Files are sharded Parquet. Both runs retain the original persona IDs.

Source and attribution

The source text comes from Persona Modality by Julius Broomfield, Kartik Sharma and Srijan Kumar. See the paper and the accompanying LICENSE-source for the MIT notice covering the upstream material.

DREAM-inspired enhancements

The default configuration adds dream_enhanced_description for all 200 personas. Each value concatenates the five profile_narrative sections from the saved DREAM-inspired expansion, in their original order. Input was the short persona text. Expansion used OpenAI GPT-5.6 Terra with low reasoning effort. Additional appearance, demographic and narrative details are synthetic. The complete profiles, source IDs, expansion settings and review warnings are available in dream-enhanced-profiles.json.

The methodology and adapted prompts derive from DREAM, upstream revision 9391ff124fa8b6ae2b7332a735fa82c80c74d322, licensed CC BY-NC 4.0. Adaptation code is in the persona-portraits fork. The original persona dataset attribution and license remain applicable.