Apply for a GPU community grant: Personal project

#1
by likeyellow - opened

What I'm building

A demo API that serves a Korean restaurant-review sentiment model, fine-tuned
from klue/bert-base. Given only the review text, it returns a continuous
1.0–5.0 star rating plus a confidence score.

Model: https://huggingface.co/likeyellow/klue-review-star
Code: https://github.com/likeyellow/review-star-api
Demo: https://likeyellow.github.io/review-star-ai/

Why it matters

In Korea, "rating terrorism" has become a recurring news story: people leave
positive review text such as "good" or "friendly" while assigning one star,
dragging a restaurant's average down. Because the text contains nothing
abusive or false, platforms cannot moderate it. One restaurant owner shut
down; when the reviews were later removed, the rating recovered from 2.9 to
5.0 — evidence that those stars never reflected the text at all.

This project removes the star input entirely and derives the rating from the
review body, so the two can no longer diverge.

Technical approach

The source dataset (KR3) ships categorical labels rather than the original
star values, so direct regression is not possible. Instead the model performs
3-class classification (negative / positive / neutral) and computes a
continuous rating as the expected value over class anchors [1, 5, 3]. The
entropy of the same probability distribution yields a confidence score, which
is used to down-weight ambiguous reviews when aggregating a restaurant's
average.

Test set: accuracy 0.768, macro F1 0.751, star MAE 0.640.
I also ran a comparison against beomi/KcELECTRA-base and documented the
trade-off (higher minority-class recall, lower overall accuracy).

Openness

  • Dataset: KR3, publicly available on the Hub (CC BY-NC-SA 4.0)
  • Model weights: published publicly on the Hub
  • Serving code and frontend: MIT licensed on GitHub
  • Training notebook will be published alongside the model card

What I need

CPU Basic is sufficient — the model is bert-base and runs inference on CPU in
well under a second. I'm currently blocked by the CPU Basic quota limit on a
free account, even with no other Spaces running.

This is a student project (2-month AI course, final project). I'd like the
demo to stay reachable so it can serve as a public reference for the
approach.

Sign up or log in to comment