A distilbert-base-uncased student distilled from the textattack/bert-base-uncased-SST-2 teacher on GLUE SST-2 sentiment classification. Training combines temperature-scaled KL divergence against the teacher's soft labels with the task cross-entropy, using a linear alpha curriculum (0.9 -> 0.1) that starts distillation-heavy and anneals toward the true labels.

Results

Single small-scale run on an RTX 5090: SST-2, 2000 train / 500 val, 1 epoch.

metric	value
teacher accuracy	0.9340
student accuracy	0.8600
accuracy retained (student / teacher)	0.9208 (92.1%)
student speedup over teacher	2.12x
param compression (teacher / student)	1.64x
student params	66,955,010

The distilled student keeps 92% of the teacher's SST-2 accuracy while running 2.1x faster with 1.64x fewer parameters. These are small-scale (1-epoch, 2000-example) numbers meant to illustrate the accuracy-for-speed trade, not a full-scale benchmark.

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

name = "narinzar/teacher-student-distillation-pipeline"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForSequenceClassification.from_pretrained(name)

inputs = tok("a thoughtful, moving film", return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
print(logits.argmax(-1).item())  # 0 = negative, 1 = positive

Labels

0: negative
1: positive

Training code: https://github.com/narinzar/teacher-student-distillation-pipeline