Soften verbiage: posters are repository-labeled, not individually verified

README.md

@@ -32,24 +32,27 @@ Developed by the [**FAIR Data Innovations Hub**](https://fairdataihub.org/) at t
 
 Text extracted from **real scientific poster PDFs** and **real non-poster documents** — zero synthetic data. Every sample comes from an actual PDF downloaded from Zenodo or Figshare as part of the posters.science corpus.
 
+This is a **balanced** dataset: 1,803 poster samples and 1,803 non-poster samples, drawn from the source corpus described below.
+
 ### Source Corpus
 
-Sampled from a curated collection of **30,000+ classified scientific PDFs**:
+Sampled from a collection of **30,000+ scientific PDFs** scraped from Zenodo and Figshare:
 
-| Category | Count | Platforms |
-|----------|-------|-----------|
-| Verified scientific posters | 28,111 | Zenodo, Figshare |
-| Verified non-posters | 2,036 | Zenodo, Figshare |
+| Category | Count | Selection Method |
+|----------|-------|-----------------|
+| Repository-labeled posters | ~28,000 | Records tagged as "poster" in Zenodo/Figshare metadata |
+| Manually confirmed non-posters | 2,036 | Flagged by structural classifier, then human-reviewed |
 | Corrupt/unreadable | 58 | — |
-| **Total classified** | **30,205** | — |
 
 Non-posters include multi-page papers, conference proceedings, abstract books, newsletters, project proposals, and other documents mislabeled as "posters" in repository metadata.
 
+**Note on poster labels**: The poster class is drawn from repository records self-described as posters by their uploaders. These were not individually verified by human reviewers. When PosterSentry was later applied to the full 30K corpus, approximately 20% of repository-labeled "posters" were reclassified as non-posters, suggesting meaningful label noise in the broader corpus. The balanced training subset published here was randomly sampled from the repository-labeled poster pool.
+
 ## Files
 
 | File | Description | Samples |
 |------|-------------|---------|
-| `poster_sentry_train.ndjson` | Training data (text + labels) | 3,606 |
+| `poster_sentry_train.ndjson` | Balanced training data (text + labels) | 3,606 |
 
 ## Format
 
@@ -64,18 +67,17 @@ NDJSON (newline-delimited JSON) with `text` and `label` fields:
 
 | Label | Count | Description |
 |-------|-------|-------------|
-| `poster` | 1,803 | Text from first page of verified single-page scientific posters |
-| `non_poster` | 1,803 | Text from first page of verified multi-page documents |
+| `poster` | 1,803 | Text from first page of repository-labeled single-page scientific posters |
+| `non_poster` | 1,803 | Text from first page of manually confirmed non-poster documents |
 
 Classes are perfectly balanced (1:1 ratio).
 
 ## Data Collection Methodology
 
-1. **Poster corpus assembly**: 30K+ PDFs scraped from Zenodo and Figshare using the [poster-repo-scraper](https://github.com/fairdataihub/poster-repo-scraper)
-2. **Classification**: A Gradient Boosting classifier using PDF structural features (page count, physical dimensions, file size) separated posters from non-posters with F1 = 1.0 on held-out data
-3. **Separation**: 2,036 non-posters moved to a separate directory; 28,111 verified posters retained
-4. **Text extraction**: First page text extracted from each PDF using PyMuPDF (fitz), cleaned and truncated to 4,000 characters
-5. **Balanced sampling**: 1,803 samples per class (limited by the smaller non-poster class)
+1. **Corpus assembly**: 30K+ PDFs scraped from Zenodo and Figshare using the [poster-repo-scraper](https://github.com/fairdataihub/poster-repo-scraper), selecting records whose metadata indicated "poster"
+2. **Non-poster identification**: A structural classifier using PDF features (page count, dimensions, file size) flagged 2,036 candidate non-posters, which were then manually reviewed and confirmed
+3. **Text extraction**: First-page text extracted from each PDF using PyMuPDF, cleaned (whitespace normalization) and truncated to 4,000 characters
+4. **Balanced sampling**: 1,803 samples randomly drawn from each class (limited by the smaller non-poster pool after feature extraction filtering)
 
 ## Related Resources
 
@@ -114,7 +116,7 @@ The poster texts in this dataset are also used by [PubGuard](https://huggingface
 
 ```bibtex
 @dataset{poster_sentry_data_2026,
-  title = {PosterSentry Training Data: Real Scientific Poster Text Corpus},
+  title = {PosterSentry Training Data: Scientific Poster Text Corpus},
   author = {O'Neill, James and Soundarajan, Sanjay and Portillo, Dorian and Patel, Bhavesh},
   year = {2026},
   url = {https://huggingface.co/datasets/fairdataihub/poster-sentry-training-data},