---
title: SlideChain Slides Dataset
url: https://www.emergentmind.com/topics/slidechain-slides-dataset
type: topic
---

# SlideChain Slides Dataset

The SlideChain Slides Dataset is a semantically annotated, blockchain-anchored collection of 1,117 high-resolution medical imaging lecture slides from a university course, designed to enable verifiable multimodal semantic extraction, reproducibility, and auditability for educational AI systems. Each slide image is paired with a transcript snippet and systematically processed using four state-of-the-art vision–language models (VLMs), producing structured concepts and relational triples documented in cryptographically hashed off-chain JSON records. The dataset serves as a foundation for benchmarking semantic disagreement, model variability, provenance research, and trustworthy pipeline development in STEM instructional contexts [2512.21684].

## 1. Dataset Composition and Organization

The corpus encompasses 23 full-length university lectures in medical imaging, covering x-ray physics, computed tomography (CT) reconstruction, magnetic resonance imaging (MRI) principles, ultrasound imaging, PET/SPECT, and multidimensional signal processing. Slide images are provided in JPEG or PNG format at standard university resolution (typically 1920×1080 or higher), with each slide uniquely identified by a (lecture_id, slide_id) tuple.

- Total slides: 1,117
- Directory structure: Separate folders for each lecture containing subfolders for images, transcript snippets, and provenance JSONs.
- File naming convention: Within each lecture, slides are labeled Slide1.jpg, Slide2.jpg, … and matched to Slide1.txt and Slide1.json.

| Lecture Topics             | # Lectures | # Slides (total) |
|---------------------------|:----------:|:----------------:|
| X-ray physics             | 23         | 1,117            |
| CT, MRI, PET/SPECT, Ultrasound, Multidimensional Signal Processing | — | — |

This organization ensures precise, reproducible referencing for semantic extraction, model comparison, and provenance anchoring.

## 2. Data Modalities and Provenance Formats

Each slide is associated with three artifacts:
- **Image**: High-resolution JPEG/PNG
- **Transcript**: Plain-text snippet (.txt)
- **Provenance record**: JSON file (.json) per slide, documenting multimodel semantic annotations and metadata

Sample directory layout:
```
Lectures/
 ├─ Lecture 1/
 │    ├─ Images/Slide1.jpg
 │    ├─ Texts/Slide1.txt
 │    └─ Provs/Slide1.json
 └─ …
```

The provenance JSON adopts a deterministic schema with sorted keys to guarantee reproducible cryptographic hashing. Each file records:
- Lecture and slide identifiers
- Model outputs (concepts, relational triples, evidence, raw output)
- Metadata (timestamps, pipeline source, hash input format)
- Paths referencing image, text, and JSON files

## 3. Semantic Annotation and Provenance Structure

Semantic annotations extracted by each VLM are organized as follows:

- **Concepts**: Flat lists of (category, term) pairs; categories include modality, anatomy, workflow, physics, software, etc.
- **Relational triples**: Structured as (subject $s$, predicate $p$, object $o$) with optional confidence scores to capture inter-concept relationships.
- **Evidence**: Quoted textual snippets or model-generated explanations supporting each annotation.

Example JSON annotation:
```json
{
  "lecture": "Lecture 1",
  "slide_id": 1,
  "models": {
    "InternVL3-14B": {
      "concepts": [ {"category":"modality","term":"medical imaging"}, … ],
      "triples": [ {"s":"Medical imaging","p":"uses","o":"science and engineering","confidence":1.0}, … ],
      "evidence": [ … ],
      "raw_output": "…"
    },
    // Qwen2-VL-7B, Qwen3-VL-4B, LLaVA-OneVision
  },
  "paths": {
    "image":"…/Lecture1/Images/Slide1.jpg",
    "text":"…/Lecture1/Texts/Slide1.txt",
    "json":"…/Lecture1/Provs/Slide1.json"
  },
  "metadata": {
    "timestamp":"2024-04-10T15:23:45Z",
    "source":"SlideChain v1.0 pipeline",
    "hash_input_format":"lexicographically sorted JSON"
  }
}
```
All annotation keys are serialized in sorted order for deterministic hashing using Keccak-256.

## 4. Blockchain Registration and Tamper-Evident Provenance

The integrity of each semantic annotation is cryptographically anchored on a local EVM-compatible blockchain using a minimal Solidity smart contract. The hash for each slide is computed via
$$
h_{L,s} = \text{keccak256}(\text{UTF-8 bytes of sorted JSON}(L, s))
$$
and registered using:
- **registerSlide**$(lectureId, slideId, slideHash, uri)$
- **getSlide**$(lectureId, slideId) \rightarrow$ SlideRecord
- **isRegistered**$(lectureId, slideId) \rightarrow$ bool

Smart contract pseudocode:

```solidity
struct SlideRecord {
  uint256 lectureId;
  uint256 slideId;
  string  slideHash;    // 0x… keccak256
  string  uri;          // off-chain JSON path
  uint256 timestamp;    // block.timestamp
  address registrant;
}
mapping(bytes32 => SlideRecord) records;  // key = keccak256(lectureId, slideId)
```

This framework supports tamper-evident, persistent semantic baselines, enabling deterministic retrieval and verification of extracted annotations.

## 5. Annotation Methodology and Quality Control

Annotations are generated by four vision–language models:
- InternVL3-14B
- Qwen2-VL-7B
- Qwen3-VL-4B
- LLaVA-OneVision (backed by Qwen2-7B)

The four-step pipeline per (slide, model) consists of:
1. Prompt construction (slide image + transcript snippet)
2. Model inference producing lists, dicts, free text outputs
3. Normalization: null-safe parsing, flattening, canonical triple extraction, lowercasing, whitespace normalization, deduplication
4. Deterministic JSON generation with key sorting for cryptographic hashing

Inter-model agreement is evaluated using:
- Concept disagreement: $D_\text{concept}(s) = |\bigcup_i C_i(s)|$
- Triple disagreement: $D_\text{triple}(s) = |\bigcup_i T_i(s)|$
- Pairwise Jaccard index: $J(C_a, C_b) = |C_a \cap C_b| / |C_a \cup C_b|$ (analogous for triples $T$)

Quality control methods include independent extraction runs yielding bit-identical JSON files (Jaccard=1.0 for every slide/model) and tamper injection testing on random JSONs, resulting in perfect detection of modifications.

## 6. Accessibility, Usage Practices, and Downstream Applications

No official train/validation/test splits are provided, reflecting the dataset’s primary purpose for semantic extraction, audit, and reproducibility, rather than supervised learning or model training.

- **Access**: Fully available via public GitHub repository (subject to slide text/image licensing restrictions).
- **Licensing**: MIT License for code and JSON artifacts; slide image/text governed by the copyright policy of contributing instructors.
- **Citation**: Manik, M. M. H., Islam, M. Z., & Wang, G. (2024), "SlideChain: Semantic Provenance for Lecture Understanding via Blockchain Registration" [2512.21684].

Potential downstream applications include:
- VLM benchmarking through quantitative disagreement and Jaccard analysis
- Identification of slides with high cross-model disagreement for targeted review or improved prompt design
- Long-term studies of semantic drift and reproducibility across model/prompt evolution
- Tamper-proof documentation for compliance in medical and legal education domains
- Classification of semantic stability (stable/moderate/unstable) using concept disagreement thresholds
- Automation of integrity checks via on-chain verification scripts and dashboards visualizing agreement metrics

## 7. Context and Research Significance

The SlideChain Slides Dataset represents the first systematic framework for constructing, tracking, and auditing multimodal semantic knowledge extracted from STEM lecture slides using blockchain-backed provenance. The pronounced cross-model discrepancies in concept overlap and relational triple agreement documented in the dataset suggest that robust semantic verification is critical in high-stakes educational and professional contexts. The reproducibility and tamper-evidence features provide a foundation for trustworthy instructional pipelines and semantic baselines as AI systems evolve. A plausible implication is the enablement of persistent audit trails and compliance monitoring for AI-generated content in domains subject to regulatory or pedagogical scrutiny.

Source: https://www.emergentmind.com/topics/slidechain-slides-dataset