Skip to content

BMED 2300 benchmark corpus and evaluation harness

Product or Programme · All products and programmes
Evidence · no claims cited

Description

A benchmark dataset and evaluation harness developed by Md Zabirul Islam et al. containing course lectures, slides, aligned transcripts, and automated metrics for evaluating pedagogical video generation.

Components

  • BMED 2300 benchmark corpus and evaluation harness for course-grounded pedagogical video generation: A reusable benchmark artifact built from the BMED 2300 Bio-Imaging course, packaged with "twenty-three lectures, 1,116 slides, aligned transcripts, a multi-repetition LLM-judge rubric, and regex-grounded objective metrics." Chunks are slide-granular, each embedding preserving a one-to-one link between a textual explanation and its slide image, stored in a flat inner-product FAISS index with 3,072-dimensional vectors. The harness pairs repeated LLM-judge rubric scoring (1-5 scale, n=3 reps) with deterministic metrics such as Flesch reading ease and regex counts of analogies and retrieval prompts. (Md Zabirul Islam et al. (2026))

Claims

Key Sources

  • Md Zabirul Islam, Md Motaleb Hossen Manik, and Ge Wang. (2026). CourseBlueprint: A Structured Pipeline for Adaptive Pedagogical Video Generation Grounded in Course Corpora. arXiv preprint. https://arxiv.org/abs/2606.20608
Page metadata
Field Value
type product
id bmed-2300-benchmark-corpus-and-evaluation-harness
title BMED 2300 benchmark corpus and evaluation harness
description A benchmark dataset and evaluation harness developed by Md Zabirul Islam et al.
status review
product_kind dataset
generated by process:settle-candidates on 2026-10-10
trust unverified — no one has confirmed this page's content

Frontmatter is the page's machine-readable record — see the Schema & Guide.