NDE Narratives Analysis Demo

What This Demo Does

This public demo is a lightweight, stage-based version of the NDE analysis workflow.

  • Goal: map narrative text into structured section-level signals.
  • Input requirement: all three sections are required (context, experience, aftereffects).
  • Model backend: default local Qwen 3.5 0.8B + optional routed HF provider models.
  • Scope: this is a public educational demo, not the full research pipeline.

Why This Matters

NDEs can be represented in two ways that do not always match perfectly:

  • free narratives (rich, contextual, temporal), and
  • questionnaire-style labels (structured and standardized).

This demo helps you inspect that representational gap by showing the extraction process step by step.

Core Questions

  1. How much can narrative-derived signals align with questionnaire-style labels?
  2. Is narrative tone sufficient to recover self-reported valence?
  3. Are experiential-perceptual features easier to recover than reflective aftereffects?
  4. Does disagreement indicate model error, representational mismatch, or both?

What This Demo Reproduces

  • Section-level analysis of context, core experience, and aftereffects.
  • LLM extraction of tone, contextual framing, and structured features.
  • Evidence-focused outputs for transparent interpretation.
  • Optional valence comparison against overall experience-weighted tone.

Demo Purpose

This page demonstrates how an LLM can extract structured information from narrative text.

Use it to compare:

  • what the model extracts, and
  • what you would extract manually from the same narrative.

Quick User Guide

How To Use This Demo

  1. Paste text into all three required sections.
  2. (Optional) add a valence label for alignment checking.
  3. Click Run Analysis.
  4. Review stage outputs and compare model extraction with your own reading.

Model Scope in This Public Space

  • Default model is Qwen 3.5 0.8B local download (runs in this Space instance after first load).
  • Additional routed models are listed only when providers are live for conversational inference.
  • Routed usage depends on your Hugging Face account quota and provider availability.
  • Prompt templates are aligned with the main local workflow.
  • It is useful for fast public testing, but it is not equivalent to the full local research pipeline.
  • For the more solid and reproducible local workflow (with Ollama models and full experiment tooling), use: https://github.com/cristian-pulido/NDE_NARRATIVES_ANALYSIS

Results Overview Video

This short video summarizes the representational-mismatch perspective and main findings.
0:00 / 0:00

Important Disclaimer

  • This tool is for research and educational use only; it is not a medical or psychological diagnostic system.
  • Do not submit personally identifying or highly sensitive information.
  • Model outputs may be inaccurate, incomplete, or biased and require human review.

Stage 1 — Input
Model

Default option runs locally in this instance; other options use routed HF providers when available.
Optional Valence

Only used for valence alignment check.

Stage 2 — Segmentation

Segmented Narrative

Stage 3 — Module Analysis

Section-Level Summary

Stage 4 — Interpretation