Release 1.0.0: human-validated labels

3,381 documents with labels from the three-reviewer survey and blinded adjudication; rows now carry document id, DOI, source, and label provenance. Supersedes the heuristic-label release.

README.md

@@ -24,60 +24,63 @@ task_categories:
 
 # PosterSentry Training Data
 
-Training dataset for [**PosterSentry**](https://huggingface.co/fairdataihub/poster-sentry) — the multimodal scientific poster classifier used in the [posters.science](https://posters.science) quality control pipeline.
+Human-validated training dataset for [**PosterSentry**](https://huggingface.co/fairdataihub/poster-sentry), the multimodal scientific poster classifier used in the [posters.science](https://posters.science) quality control pipeline.
 
 Developed by the [**FAIR Data Innovations Hub**](https://fairdataihub.org/) at the California Medical Innovations Institute (CalMI²).
 
-## Dataset Description
+## Version
+
+| Version | Date | Notes |
+|---------|------|-------|
+| **1.0.0** | 2026-08-18 | Human-validated labels: every document was independently rated by three reviewers (Krippendorff's alpha 0.79) and the 439 contested documents were settled in a blinded adjudication review. Rows now carry document identifiers, DOIs, and label provenance. This is the corpus the published PosterSentry model is trained on. |
 
-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.
+The earlier unversioned release (April 2026) used heuristic repository labels and is superseded; it remains available in the repository history.
 
-This is a **balanced** dataset: 1,803 poster samples and 1,803 non-poster samples, drawn from the source corpus described below.
+## Dataset Description
 
-### Source Corpus
+Text extracted from **real scientific documents**, zero synthetic data. Every sample comes from an actual PDF downloaded from Zenodo or Figshare as part of the posters.science corpus, and every label was validated by people.
 
-Sampled from a collection of **30,000+ scientific PDFs** scraped from Zenodo and Figshare:
+The corpus holds **3,381 documents**: 1,686 posters and 1,695 non-posters (near-balanced by construction of the candidate pools, not by resampling).
 
-| 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 | — |
+### Labeling
 
-Non-posters include multi-page papers, conference proceedings, abstract books, newsletters, project proposals, and other documents mislabeled as "posters" in repository metadata.
+1. **Survey**: Three reviewers independently classified all 3,570 candidate documents (poster, non-poster, or unsure) at [survey.posters.science](https://survey.posters.science), casting 10,710 votes with an inter-rater Krippendorff's alpha of 0.79.
+2. **Adjudication**: The 439 documents without a unanimous panel (377 decided two to one, 62 exact ties) were adjudicated in a blinded review: each document image was re-examined in randomized order without access to the panel votes or any model output.
+3. **Deduplication**: 182 near-duplicate documents (matching normalized 300-character text prefixes) and 7 documents with unavailable PDFs were removed, leaving the 3,381-document training corpus published here.
 
-**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.
+Non-posters include multi-page papers, conference proceedings, abstract books, newsletters, flyers, slide decks, and other documents mislabeled as "posters" in repository metadata.
 
 ## Files
 
-| File | Description | Samples |
-|------|-------------|---------|
-| `poster_sentry_train.ndjson` | Balanced training data (text + labels) | 3,606 |
+| File | Description | Rows |
+|------|-------------|------|
+| `poster_sentry_train.ndjson` | Human-validated training corpus (text, labels, identifiers) | 3,381 |
 
 ## Format
 
-NDJSON (newline-delimited JSON) with `text` and `label` fields:
+NDJSON (newline-delimited JSON), one document per row:
 
 ```json
-{"text": "TITLE: Effects of Temperature on Enzyme Kinetics\nAUTHORS: A. Smith...", "label": "poster"}
-{"text": "Abstract. We present a novel approach to distributed computing...", "label": "non_poster"}
+{"id": "fxd4ylwf0byrtj307b5k3kpm", "doi": "10.5281/zenodo.1234567", "source": "zenodo", "text": "TITLE: Effects of Temperature on Enzyme Kinetics ...", "label": "poster", "label_source": "unanimous_panel"}
 ```
 
-## Label Distribution
-
-| Label | Count | Description |
-|-------|-------|-------------|
-| `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 |
+| Field | Description |
+|-------|-------------|
+| `id` | Survey document identifier (matches the paper's supplementary files) |
+| `doi` | DOI of the source repository record (Zenodo or Figshare) |
+| `source` | `zenodo` or `figshare` |
+| `text` | First-page text extracted with PyMuPDF, whitespace-normalized, truncated to 4,000 characters |
+| `label` | `poster` or `non_poster` (human-validated) |
+| `label_source` | `unanimous_panel` (all non-unsure votes identical) or `adjudicated` (settled in the blinded review) |
 
-Classes are perfectly balanced (1:1 ratio).
+## Label Distribution
 
-## Data Collection Methodology
+| Label | Count |
+|-------|-------|
+| `poster` | 1,686 |
+| `non_poster` | 1,695 |
 
-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)
+By provenance: 2,949 documents carry a unanimous panel label and 432 an adjudicated label (7 of the 439 adjudicated documents fell to deduplication or unavailable PDFs). Every row carries a DOI.
 
 ## Related Resources
 
@@ -86,6 +89,7 @@ Classes are perfectly balanced (1:1 ratio).
 | **PosterSentry model** | [fairdataihub/poster-sentry](https://huggingface.co/fairdataihub/poster-sentry) |
 | **poster-sentry** | [GitHub](https://github.com/fairdataihub/poster-sentry) |
 | **poster-sentry-training** | [GitHub](https://github.com/fairdataihub/poster-sentry-training) |
+| **poster-sentry-evaluation-paper-code** | [GitHub](https://github.com/fairdataihub/poster-sentry-evaluation-paper-code) |
 | **Llama-3.1-8B-Poster-Extraction** | [fairdataihub/Llama-3.1-8B-Poster-Extraction](https://huggingface.co/fairdataihub/Llama-3.1-8B-Poster-Extraction) |
 | **poster2json library** | [PyPI](https://pypi.org/project/poster2json/) · [GitHub](https://github.com/fairdataihub/poster2json) |
 | **poster-json-schema** | [GitHub](https://github.com/fairdataihub/poster-json-schema) |
@@ -93,13 +97,6 @@ Classes are perfectly balanced (1:1 ratio).
 
 ## Usage
 
-### Train PosterSentry from this data
-
-```bash
-pip install poster-sentry
-python scripts/train_poster_sentry.py --n-per-class 2000
-```
-
 ### Load directly with HuggingFace datasets
 
 ```python
@@ -107,7 +104,7 @@ from datasets import load_dataset
 
 ds = load_dataset("fairdataihub/poster-sentry-training-data")
 print(ds["train"][0])
-# {"text": "TITLE: ...", "label": "poster"}
+# {"id": "...", "doi": "...", "source": "zenodo", "text": "TITLE: ...", "label": "poster", "label_source": "unanimous_panel"}
 ```
 
 ### Use for PubGuard doc_type training
@@ -119,8 +116,9 @@ The poster texts in this dataset are also used by [PubGuard](https://huggingface
 ```bibtex
 @dataset{poster_sentry_data_2026,
   title = {PosterSentry Training Data: Scientific Poster Text Corpus},
-  author = {O'Neill, James and Soundarajan, Sanjay and Portillo, Dorian and Patel, Bhavesh},
+  author = {O'Neill, Jamey and Portillo, Dorian and Zeinali, Nahid and Soundarajan, Sanjay and Blake, Gerard and Sarin, Parth and Buttrick, Adam and Patel, Bhavesh},
   year = {2026},
+  version = {1.0.0},
   url = {https://huggingface.co/datasets/fairdataihub/poster-sentry-training-data},
   note = {Part of the posters.science initiative}
 }
@@ -128,11 +126,11 @@ The poster texts in this dataset are also used by [PubGuard](https://huggingface
 
 ## License
 
-MIT License — See [LICENSE](https://opensource.org/licenses/MIT) for details.
+MIT License. See [LICENSE](https://opensource.org/licenses/MIT) for details.
 
 ## Acknowledgments
 
 - [FAIR Data Innovations Hub](https://fairdataihub.org/) at California Medical Innovations Institute (CalMI²)
 - [posters.science](https://posters.science) platform
 - HuggingFace for dataset hosting infrastructure
-- Funded by The Navigation Fund ([10.71707/rk36-9x79](https://doi.org/10.71707/rk36-9x79)) — "Poster Sharing and Discovery Made Easy"
+- Funded by The Navigation Fund ([10.71707/rk36-9x79](https://doi.org/10.71707/rk36-9x79)), "Poster Sharing and Discovery Made Easy"