# app.py – DivPol Creativity Study (Hugging Face Space)
# ---------------------------------------------------------
# UI: Original two-panel layout (GPT chat left, scoring right)
# Scoring: 3 embedding models × pool distance → z-score → Φ(z) → average × 100
# - New responses are chunked (~300 chars each)
# - Each chunk is scored against the AI reference pool (pre-parsed chunks)
# - Chunk percentiles are averaged within each model, then averaged across models
# Flow: Prolific ID → 3 tasks (randomised) × 5 submissions → Qualtrics redirect
# ---------------------------------------------------------
#
# CHANGELOG (v4 – Feb 2026)
# --------------------------
# [INSTR] New welcome + instruction text per participant script
# [CONT] "Continue" button between tasks (manual advance)
# [TERM] "Prompt X/3" → "Task X/3"; "Round/Attempt" → "Submission"
# [LABEL] Score bar: "divergence" → "distinctiveness"
# [CHAR] Submission blocked outside 300–600 chars; live counter in status box
# [TABLE] mpnet/noinstruct/gist columns removed; "DivPol" → "Distinctiveness Score"
# [REDIR] Qualtrics redirect on study completion (placeholder URL)
# ---------------------------------------------------------
import os, json, random, hashlib, threading, csv as _csv, re
from pathlib import Path
from typing import List, Dict, Any
from datetime import datetime, timezone
import numpy as np
import pandas as pd
import gradio as gr
from scipy.stats import norm
import torch
from transformers import AutoTokenizer, AutoModel
try:
from openai import OpenAI
_HAS_OPENAI = True
except Exception:
_HAS_OPENAI = False
# ============================================================
# Config
# ============================================================
QUALTRICS_REDIRECT_URL = "https://georgetown.az1.qualtrics.com/jfe/form/SV_9FaybvTxkZOaVFk"
CHAR_MIN = 300
CHAR_MAX = 600
SUBMISSIONS_PER_TASK = 5
# Column list for the history table
HIST_COLUMNS = ["Submission", "Response Preview", "Distinctiveness Score"]
OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-4o-mini")
EMB_MODELS = [
"sentence-transformers/all-mpnet-base-v2",
"avsolatorio/NoInstruct-small-Embedding-v0",
"avsolatorio/GIST-Embedding-v0",
]
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ============================================================
# Paths
# ============================================================
BASE_DIR = Path(__file__).resolve().parent
PERSISTENT_DIR = Path("/data")
if PERSISTENT_DIR.exists():
CACHE_DIR = PERSISTENT_DIR / "cache"
DATA_DIR = PERSISTENT_DIR / "responses"
os.environ["HF_HOME"] = str(PERSISTENT_DIR / ".huggingface")
else:
CACHE_DIR = BASE_DIR / "cache"
DATA_DIR = BASE_DIR / "data"
CACHE_DIR.mkdir(exist_ok=True)
DATA_DIR.mkdir(exist_ok=True)
# ============================================================
# Tasks
# ============================================================
PROMPTS = {
"car": {
"name": "Car Safety Feature",
"text": (
"Create a new feature for a car that would help keep "
"drivers and pedestrians safe."
),
"ref_file": "car_xgb_reference_plus_responses.xlsx",
},
"teambuilding": {
"name": "Team Building Activity",
"text": (
"What are some ways to do teambuilding on video conferencing, "
"with each person only needing a piece of paper and a rubber band?"
),
"ref_file": "teambuilding_xgb_reference_plus_responses.xlsx",
},
"routine": {
"name": "Morning Routine",
"text": (
"Design a 20-minute morning routine that helps someone who "
"wants to start their day in a better mindset."
),
"ref_file": "routine_xgb_reference_plus_responses.xlsx",
},
}
# ============================================================
# Instruction text [INSTR]
# ============================================================
WELCOME_TEXT = """\
👋 Welcome! To start, please enter your Prolific ID and click the "Start Study" button.
"""
INSTRUCTION_TEXT = """\
On the left panel, you will work with an AI chatbot. You may send messages \
to the chatbot by typing and clicking the "Send" button, and you will receive \
interactive responses. You may interact with it as many or as few times as you like \
(but please use it at least once per task).
\
On the right panel, the "Sketchpad" allows you to draft and organize your response \
(please aim for 300–600 characters. Your submission will be blocked otherwise). \
When you're ready, click "Copy to Submission Box" to move your response to the \
submission box, then click "Submit and Score."
\
You will complete five submissions per task, for a total of three tasks. \
After each submission, you will receive a Distinctiveness Score out of 100:
\
• A score closer to 0 means your response is very similar \
to a typical AI-generated response.
\
• A score of 50 indicates your response is moderately similar \
to a typical AI-generated response.
\
• A score closer to 100 means your response is very different \
from a typical AI-generated response.
\
Your goal is to refine and develop your ideas so your submissions become increasingly \
distinct from the AI response. You will be able to see your response history and scores \
in a table after each submission, for a total of five submissions per task.
\
Once you finish one task, click the "Continue to the Next Task" button to move on to the next.\
"""
def make_prompt_html(prompt_cfg):
return (
f''
f'
📝 {prompt_cfg["name"]}
'
f'
{prompt_cfg["text"]}
'
f'
{INSTRUCTION_TEXT}
'
f'
0 – Very similar to AI
100 – Very different from AI
{pct:.1f}
{label}
"""
# ============================================================
# Qualtrics redirect HTML
# ============================================================
def make_redirect_html(prolific_id):
url = f"{QUALTRICS_REDIRECT_URL}?PROLIFIC_PID={prolific_id}"
return (
f''
f'
Study complete — thank you!
'
f'
You will be redirected to the survey '
f'shortly. If not, '
f'click here.
'
f''
f'