datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
glam-extraction-benchmark
GLAM extraction benchmark
Structured extraction from cultural-heritage documents. The first configuration is
nls-index-cards: 98 manuscript catalogue cards from the National Library of Scotland.
Source and credits
Derived from NationalLibraryOfScotland/index-cards-eval,
revision 2a81070549d8493c2c538744a9dbbc1dc72cb146 (CC0). Images and checked outputs are preserved.
NLS cataloguers reviewed the model-drafted labels: 66 accepted as drafted, 32 corrected.
Drafting… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/glam-extraction-benchmark.bl-crop-tighten-v1
bl-crop-tighten-v1
Training data for crop tightening on the British Library Book Images collection: 7,565 ABBYY picture-block crops (train 6,050 / validation 757 / test 758) with instance boxes and segmentation masks. The splits are book-safe — no book appears in more than one split (4,484 books total).
The labels are weak labels, not human annotations: tiiuae/Falcon-Perception-0.6B ran open-vocabulary segmentation over 8,400 stratified crops (embellishments, plates, medium… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/bl-crop-tighten-v1.synthetic-linkedart-production-destructionindex-card-detection-v3
Dataset Card for Archival Index Card Detection — mixed collections
A training dataset for object detection of index cards in archival scans. Combines four publicly-released collections — NLS Advocates Library single-card pages, US Navy Nurse Corps multi-card biographical sheets, Boston Public Library catalog cards, and Duke Rubenstein manuscript catalog cards — into a single object-detection schema.
Dataset Details
Dataset Description
1,425 archival scans… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/index-card-detection-v3.synthetic-parsed-names-yaml
Dataset Card for Synthetic Parsed Names (YAML)
This dataset contains approximately 500,000 synthetic examples of complex, unstructured historical names paired with their structured YAML equivalents. It is designed to fine-tune small open-source large language models (LLMs) to accurately parse cultural heritage name strings into isolated components (first names, last names, middle names, dates, titles, etc.) for de-duplication and structured data ingestion.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/synthetic-parsed-names-yaml.index-card-detection-v5
Dataset Card for Archival Index Card Detection — v5 (ensemble-relabelled navy)
Refined version of small-models-for-glam/index-card-detection-v3. All NLS / BPL / Rubenstein rows are passed through unchanged. The 25 navy-nurse-corps rows have their bounding boxes re-labelled via a v3+v4 model ensemble plus human review, replacing the SAM3-only bootstrap from v3.
Dataset Details
Dataset Description
Same 1,425-row mixed-collection composition as v3. The… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/index-card-detection-v5.olmo2_delta_gpt5_vs_smallmodelssynthetic-linkedart-physical-characteristicssynthetic-aat-materials
Synthetic AAT Materials Dataset
Dataset Description
This dataset contains 1000 synthetic examples of cultural heritage object descriptions paired with their materials as they would appear in the Getty Art & Architecture Thesaurus (AAT). The data is formatted for training conversational AI models, particularly Qwen3, to identify and extract materials from cultural heritage object descriptions.
Dataset Structure
Each example contains:
messages: Conversation… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/synthetic-aat-materials.small-model-schema-gym
Small Model Schema Gym Dataset
Deterministically generated chat examples for first-pass JSON compliance on
the project Dream brief and Safety plan contracts.
Files
train.jsonl: 500 training examples.
validation.jsonl: 200 held-out examples.
manifest.json: counts, seed, provenance, and overlap check.
Each row contains:
{
"id": "stable example identifier",
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."}… See the full description on the dataset page: https://huggingface.co/datasets/KwabsHug/small-model-schema-gym.aat-real-world
Real-World AAT Materials Dataset
This dataset contains 189,523 real-world examples of cultural heritage object material descriptions paired with their corresponding Art & Architecture Thesaurus (AAT) material classifications.
Dataset Description
The dataset is designed for training models to extract material information from cultural heritage object descriptions. Each example consists of:
Input: A real material description from cultural heritage collections
Output:… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/aat-real-world.agenda-parser-models-example-agent-traces
Agenda Parser — fine-tuned agent models
Three Gemma 4 models fine-tuned to drive the Agenda Parser's ReAct agent: at each step
the model emits a single JSON action {"thought","tool","args"} over two toolkits —
meeting-agenda packets and Michigan local-government law (Open Meetings Act, FOIA,
the Michigan Compiled Laws via Cornell LII). This card doubles as the project write-up; the
dataset itself (bottom) is a gallery of example traces from the three models.
tier
base… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/agenda-parser-models-example-agent-traces.index-card-blank-content
Index-card blank / content / divider classifier — dataset
Cropped single archival index cards labelled blank, content, or divider, for
training a tiny CPU pre-filter that skips blank/divider cards before expensive VLM metadata
extraction in card-catalogue digitisation pipelines.
Two collections: Boston Public Library (BPL) FRC shelf-list cards and National Library
of Scotland (NLS) Advocates Library cards. Styles differ, so evaluate per collection.
How it was made… See the full description on the dataset page: https://huggingface.co/datasets/small-models-for-glam/index-card-blank-content.responses-and-asr-labels-small-models
LLM Responses and ASR Labels — Small Models
Model responses to harmful prompts, labelled by 4 LLM-as-judge guards.Companion dataset for the master's thesis ASR Signal Geometry: Dense Representations vs. SAE Features (HSE, 2025).
Dataset composition
N = 4 326 prompts per model, (no adversarial suffix). Two sources:
Source
N
Description
JailbreakBench ()
100
Curated harmful behaviours
Anthropic HH-RLHF red-team-attempts ()
4 226
Red-team conversations… See the full description on the dataset page: https://huggingface.co/datasets/SabrinaSadiekh/responses-and-asr-labels-small-models.synthetic-parsed-namesevalap-compare-albert-small-with-apertus-small-models-82
Compare albert-small with apertus-small models (ID: 82)
Comparing albert-small model with apertus-small model alone, and in a RAG setting with service-public + travail-emploi sheets, on a french administration Q/A datasets
Overview
This dataset contains 8 experiments
from the EvalAP evaluation platform.
Datasets: Assistant IA - QA, MFS_questions_v01
Models evaluated: meta-llama/Llama-3.1-8B-Instruct, swiss-ai/Apertus-8B-Instruct-2509
Metrics: generation_time… See the full description on the dataset page: https://huggingface.co/datasets/AgentPublic/evalap-compare-albert-small-with-apertus-small-models-82.olmo2_delta_typo_vs_og_smallmodelsolmo2_delta_gpt5_vs_smallmodels_v2small_model_samples_2