pcwoods/2026-24679-distilbert-study-spots

Study-Spot Verdict Classifier

This model classifies a short description of a campus location as Not good for studying (0) or Good for studying (1). It fine-tunes DistilBERT with norm-fit PEFT in AutoGluon MultiModal.

Inputs and outputs

Supply a pandas DataFrame with a text column. predict() returns 0 or 1.

Label	Meaning
0	Not good for studying
1	Good for studying

Data and training

Source: Study-spot passages, short descriptions of CMU and Pittsburgh study locations. Published splits were used unchanged. Augmented variants appear in training only, and no parent_id crosses splits.

Partition	Rows
Training	1,151
Validation	15 (originals)
Test	15 (originals)

Training setting	Recorded value
Framework	AutoGluon MultiModal 1.6.1
Text backbone	distilbert-base-uncased, max 128 tokens
Fine-tuning mode	norm_fit: 62,978 of 66.4 M parameters trainable (0.095%)
Optimizer	Learning rate 2e-5, layerwise decay, weight decay 0.01
Batch size and epochs	8; 5 epochs (720 steps)
Search space	None; fine-tuning mode chosen manually (bit_fit / norm_fit / full)
Checkpoint selection	Best single epoch by validation accuracy (epoch 2)
Text augmentation	Disabled
Seed	24679
Environment	Colab CPU (2 vCPU, 12.7 GB RAM, no GPU); Linux; Python 3.13.15; PyTorch 2.11.0+cpu

Evaluation

Metric	Validation	Test
Accuracy	0.667 (10/15)	0.733 (11/15)
Weighted F1	0.664	0.731

Three of the four test errors were Not-good passages predicted as Good, and the fourth was the reverse. The misclassified Not-good passages mix praise with a negative verdict, which suggests the model tracks sentiment words more than the conclusion. Each prediction moves accuracy by about 7 points. Only 0.095% of weights were updated, and validation did not improve after epoch 2.

Load and predict

from huggingface_hub import snapshot_download
from autogluon.multimodal import MultiModalPredictor
import pandas as pd
 
path = snapshot_download("pcwoods/2026-24679-distilbert-study-spots", allow_patterns="automm/*")
predictor = MultiModalPredictor.load(f"{path}/automm")
print(predictor.predict(pd.DataFrame({"text": ["Quiet library with big tables."]})))

The automm/ folder also contains requirements.txt and model_notes.json.

Licensing and provenance

The model is released under Apache-2.0, matching DistilBERT. Refer to the dataset card for the data's terms.

AI usage disclosure

Claude was utilized to help modify an existing notebook template. Additionally, Claude was used to assist with model card creation.