DistilBERT SST-2 FP16

An FP16 version of distilbert/distilbert-base-uncased-finetuned-sst-2-english, prepared for a quantization comparison project. This model performs binary sentiment classification on English text using the SST-2 label set.

Evaluation snapshot

• Evaluation: 512 examples from the seeded SST-2 validation subset
• Accuracy: 90.82%
• Change from FP32 accuracy: 0.000 percentage points
• Serialized artifact size: 133.95 MB
• Measured latency: 497.2 ms/example on CPU
• Environment: PyTorch 2.13.0+cpu, seed 42

The latency is from the project's CPU benchmark and is not a general speed claim. In that run, FP16 was slower than the FP32 baseline. See the project repository's report for the full methodology and limitations.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "SoloCode/distilbert-sst2-fp16-quantization"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

This checkpoint derives from the model linked above. Review its model card and license before reuse.