Downloads · 30 days
6
9% of all-time downloads
ypolatog/24679-hw2-movie-genre-distilbert
24679-hw2-movie-genre-distilbert is a text classification model from ypolatog. Use it when you need a label for a piece of text. The card lists the license as other.
distilbert-base-uncased fine-tuned to predict a movie's genre from its plot description, over 5 genres: Thriller, Action, Romance, Comedy, Animation. Trained on ArinRoths/movie-data, a classmate's HW1 dataset.
Downloads · 30 days
6
9% of all-time downloads
All-time downloads
68
Public
Parameters
67M
268 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors268 MB · 100%
From the Hugging Face model README
distilbert-base-uncased fine-tuned to predict a movie's genre from its plot description, over
5 genres: Thriller, Action, Romance, Comedy, Animation. Trained on
ArinRoths/movie-data, a classmate's HW1 dataset.
| Metric | Cross-validated (3-fold grouped) | Held-out test (n=15) |
|---|---|---|
| Macro F1 | 0.497 +/- 0.070 | 0.659 |
| Accuracy | 0.494 | 0.667 |
Chance is 0.200. Majority baseline on test: 0.200
(always predicts Thriller).
Learning rate 2e-05, selected by sweeping [2e-05, 3e-05, 5e-05] on validation macro F1.
Batch size 16, weight decay 0.01, warmup ratio 0.1, max sequence
length 128. Trained 4 of 8 permitted epochs
before early stopping (patience 2); the best checkpoint was reloaded.
The source dataset publishes label ids [0, 1, 3, 4, 5] —
id 2 is unused. This model remaps them to contiguous ids 0..4; the mapping is
stored in the model config and in split_manifest.json.
Coursework for CMU 24-679 Design AI; not for any real decision. Trained on a few hundred short descriptions collected by one student, most of them augmented paraphrases of a smaller original set, so it reflects that vocabulary and those films. The held-out test split is 15 descriptions — too few to judge on alone, which is why the cross-validated figure is quoted first.
The source dataset declares no license. Reuse should be cleared with its author, ArinRoths.
Seed 24679, transformers 4.57.6. metrics.json holds every number above,
sweep_results.csv the learning-rate sweep, and split_manifest.json the exact source_ids of
each split, so HW2 Part 4 can score the identical test set.