Hugging Face's logo

Spaces:
AE-Talk-bot
/
A
Running

App Files Community
A / README.md
linxinhua's picture
linxinhua
Fix auto-restart watchdog strptime: log filenames now include _<sessioncode> suffix; parse only the YYYYMMDD_HHMMSS prefix so idle check succeeds.
fd3bf73 verified
preview code
|
Raw
History Blame Contribute Delete
2.61 kB

A newer version of the Gradio SDK is available: 6.22.0

Upgrade
metadata 
title: v5 · AI Ethics HOT — Socratic FRAME
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.11.0
python_version: '3.11'
app_file: app.py
pinned: false
short_description: v5 · Socratic AI · FRAME variant

v5 · AI Ethics HOT — Socratic AI Discussion

Research-instrument chatbot for an RCT on critical-thinking transfer.

The participant:

  1. Reads an AI ethics case.
  2. Writes an initial 150-250 word argument.
  3. Discusses with a Socratic AI for ~20 minutes.
  4. Ends the session.

The Socratic AI runs on a two-agent architecture: a Discussion Manager (small fast classifier) decides what kind of move the bot should make each turn, and the Main Bot (the Socratic discussion partner) executes that move. Both agents read their prompts from markdown files in prompts/, so behaviour can be edited without touching code.

Configuration

Runtime is controlled by Space Secrets:

Secret Purpose
API_KEY OpenAI API key (used by manager + main bot + per-session essay generator)
HF_access HF token with write scope (logger uploads to AE-Talk-bot/log)
RUN_MODE test / pilot / prod (picks the dataset bucket: v5-HOT/socratic/<RUN_MODE>/)

Logs

Schema 1.1: every session writes one JSON to AE-Talk-bot/log/v5-HOT/socratic/<RUN_MODE>/<session_id>.json containing the participant's submitted argument, the per-session generated essay + discussion plan, the manager's cumulative verdict history, and per-turn records (user message, manager verdict + directive, plan-control text injected, bot reply, scratch fields, latencies).

Architecture

  • app.py: Gradio UI (single page, 2x2 layout) plus the two-agent orchestrator wiring.
  • prompts/manager_prompt.md: prompt for the Discussion Manager agent.
  • prompts/main_prompt.md: prompt for the main Socratic bot.
  • prompts/perfect_answer_prompt.md: prompt for the per-session essay + discussion plan generator.
  • core/orchestrator.py: turn_exchange helpers that compose manager + main bot.
  • core/manager_agent.py: manager LLM call, JSON verdict output.
  • core/main_agent.py: main bot streaming, scratch parsing.
  • core/plan_parsing.py: parses the 4-field discussion plan schema (id / probe / rationale / anchor).
  • core/perfect_answer.py: per-session essay + discussion plan generator.
  • core/logger.py: Schema 1.1 HF dataset logger.
  • data/case_analysis.md, data/ethics-frameworks.md: generator inputs (case material).
  • content/case_1.md: the case prompt shown to the participant.