krimits/distilbert-hotel-reviews

DistilBERT hotel-review sentiment classifier (full fine-tune)

Full fine-tune of distilbert-base-uncased for binary sentiment classification of hotel reviews (labels derived from the Booking.com review schema, not a score threshold). This checkpoint is the strongest model of the `krimits/hotel-review-nlp` project.

Metrics (frozen test set, 13,278 reviews)

Statistical comparison (exact McNemar, same frozen test set): full fine-tune > scratch-LoRA (p ≈ 5.0×10⁻⁶); scratch-LoRA ≈ Qwen2.5-0.5B QLoRA (p = 0.525). Full 10-pair benchmark: `runs/benchmark/results.json`.

Provenance

• —Weights are byte-identical to the Colab notebook 05_distilbert_full_and_lora_colab.ipynb run distilbert_legacy_full_v1 (commit 530cf9e14eea7e527a240ea470ae3a8e222c4d3f): model.safetensors sha256 f30daa44e795b7015151892ff5ec285e14a916e122902aa0617a15bf96ec54b7.
• —Known legacy limitation: the historical splits contain small normalized cross-split text overlap (train-test 170 rows); disclosed and fingerprinted in the experiment log.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tok = AutoTokenizer.from_pretrained("krimits/distilbert-hotel-reviews")
model = AutoModelForSequenceClassification.from_pretrained("krimits/distilbert-hotel-reviews")
inputs = tok("The room was spotless and the staff was wonderful.", return_tensors="pt")
print(model(**inputs).logits.argmax(-1).item())  # 1 = positive, 0 = negative

CPU latency: ~15–25 ms/review (single request); ~20.5 req/s under 20-user load (p50 360 ms under concurrency). Live demo: `krimits/hotel-review-demo`.