AdviceIT by Radit

Explainable AI investment advice and how people rely on it

References and tools

The short list of what AdviceIT directly builds on: the dataset the advisors are trained on, the handful of works behind the design of the study and the explanations, and the software and models it runs on. Where a work informed a specific part of the app, that part is named.

Dataset

Robo-advisory, risk and ethics

Provenance

Trust, reliance and algorithm aversion

Explainable AI methods

Human-centred explanation interfaces and measures

Software, models and services

WhatUsed forLicence and source
Python 3 (standard library)serve.py, the study server, and ml/fetch_ils_bench.py, tools/make_pages.pyPSF licence, python.org
NumPyTraining both advisors in ml/train_model.py (multilayer perceptron with Adam, multinomial logistic regression, temperature scaling)BSD, numpy.org
openpyxlReading the ILS-Bench Excel file in ml/fetch_ils_bench.pyMIT, openpyxl.readthedocs.io
WebLLM (MLC project)Running the language model in the browser for the conversational condition, the narrative reading and the benchmark. Ruan, C. F., and colleagues (2024). WebLLM: A high-performance in-browser LLM inference engine. arXiv:2412.15803.Apache 2.0, github.com/mlc-ai/web-llm, loaded from jsDelivr (esm.run)
Qwen 2.5 1.5B InstructDefault language model for the conversational delivery and the narrative readingApache 2.0, Alibaba Cloud, weights via Hugging Face (MLC builds)
Llama 3.2 1B InstructLighter alternative language modelLlama 3.2 Community License, Meta Platforms, weights via Hugging Face (MLC builds)
WebGPUGPU access in the browser for WebLLMW3C specification, w3.org/TR/webgpu
Mendeley DataHosting of the ILS-Bench dataset and its API used by the fetch scriptdata.mendeley.com
GitHub and GitHub PagesSource hosting and static deploymentpages.github.com
Node.js and jsdomDevelopment only: automated checks of the pages, the explanations and the study flowMIT, github.com/jsdom/jsdom
Browser standardslocalStorage for the session log, Blob and object URLs for downloads, the dialog element for the panels, ARIA for accessibilityWHATWG and W3C
WCAG 2.1 level AAContrast, keyboard operation, focus visibility, labels, never colour alonew3.org/TR/WCAG21

Everything else in AdviceIT (the label rules, the explanation modules, the study machinery, the pages) is original code written for this project, released under the MIT licence.