davanstrien/finepdfs-edu-purpose

FinePDFs-Edu document purpose — proof of concept

This dataset is a proof of concept for using agents to build small classifiers for data curation, then applying them cheaply with Hugging Face Jobs.

It contains a 1% sample of the English portion of FinePDFs-Edu, enriched with document-purpose predictions from a SetFit classifier. The aim is to make it easier to explore different mixtures of documents when sampling data.

Companion blog post: Using agents to build efficient classifiers for data curation.

Trained classifier and prediction helper · Original inference script.

The dataset contains 230,234 rows with their original source text. 191,724 have predictions; 38,510 failed the inference eligibility checks and retain empty predictions alongside their exclusion reasons.

Each classified document has one predicted primary purpose:

Label	Examples
exercise_assessment	Worksheets, exams, practice problems and answer keys
instruction_reference	Teaching material, explanations, practical guides and reference resources
research_analysis	Research papers, investigations and analytical essays
administration_policy	Organisational rules, requirements, schedules and business records
news_promotion	Newsletters, announcements, recruitment and promotional material
other	Other identifiable purposes, including creative expression, personal reflection and ritual texts

These categories describe document function. They do not measure educational quality, factual reliability or relevance to a particular retrieval task.

The classifier was trained using 200 agent-labelled examples, following a human-reviewed annotation guide. It uses SetFit with a ModernBERT embedding model. For longer documents, classification used excerpts from the beginning, middle and end. The text column preserves the original source text, which may itself already be truncated; is_truncated records that source flag.

On 120 held-out, agent-labelled documents, the selected model achieved 65.8% accuracy and 0.556 macro-F1. These evaluation labels are also agent-generated, so this is preliminary evidence. The other category is particularly weak. scores contains the six class probabilities, while top_score is their maximum; these scores have not been calibrated as probabilities of correctness.

Inference ran on Hugging Face Jobs and processed the sample in approximately 42 minutes, costing around $0.70 in GPU running time. Training and model-selection experiments separately used approximately $2.90 of GPU compute. Agent costs and storage are excluded.

For comparison, we estimate $13–26 to label the same documents using low-cost batch LLM APIs: approximately $13 with Gemini 2.5 Flash-Lite or $26 with GPT-5.6 Luna. This assumes an average of 1,300 input tokens per document, including the excerpt and instructions, and 10 output tokens, with thinking disabled and no caching. These are estimates using prices checked on 10 September 2026, rather than measured API runs; label-quality parity has not been tested. Google pricing, OpenAI pricing.

You could use these predictions to inspect particular document types or experiment with different mixtures for retrieval training or pretraining. We have not tested whether those mixtures improve downstream models.

For example, inspect a few predicted exercises using Datasets streaming:

from datasets import load_dataset

dataset = load_dataset(
    "davanstrien/finepdfs-edu-purpose",
    split="train",
    streaming=True,
)

exercises = dataset.filter(
    lambda row: row["eligible"]
    and row["predicted_purpose"] == "exercise_assessment"
)

for row in exercises.take(3):
    print(row["text"][:500])

Sampling covered all 100 English source files, with quotas proportional to their row counts. Randomly selected Parquet row groups supplied the sampled documents, so this is a clustered sample. Predicted category frequencies should not be treated as established proportions for the full corpus.

The dataset retains source attribution under ODC-BY. Source revisions, sampling details and checksums are recorded in the accompanying manifest. Source IDs, URLs and row positions support tracing predictions back to the original data. The train split is a loading convention; its labels are model predictions.