SweepChampion
SweepChampion is a product-review classifier fine-tuned from distilbert-base-uncased.
This release corresponds to the winning run of our internal hyperparameter sweep
(run_20240902_b7e2), selected under the production latency budget.
Training Details
Training Hyperparameters
| Hyperparameter | Value |
|---|---|
| batch_size | 32 |
| epochs | 4 |
| weight_decay | 0.02 |
| warmup_ratio | 0.1 |
| max_seq_length | 256 |
| seed | 1337 |
Training Data
Fine-tuned on the train split of the internal product-reviews-v2 dataset.
Model selection used the validation split shared by all runs of the sweep.
Evaluation Results
Final metrics of the selected run (validation split):
| Metric | Value |
|---|---|
| train_f1 | 0.9456 |
| val_f1 | 0.9127 |
| val_precision | 0.9081 |
| val_recall | 0.9174 |
| p95_latency_ms | 231.8 |
| params_m | 66.36 |
Selection Rule
Among the completed runs of the sweep, only runs whose measured p95 latency respected the production budget were eligible; runs without a latency measurement were treated as ineligible. The released run is the eligible run with the highest val_f1.
Usage
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained("SweepChampion-Release")
tokenizer = AutoTokenizer.from_pretrained("SweepChampion-Release")
Limitations
The model was trained on internal product-review data and is intended for sentiment-style classification of similar text. Latency figures were measured on our production serving hardware and may differ on other setups.
- Downloads last month
- 14