game-review-sentiment

Game Review Sentiment (int8 ONNX)

Binary sentiment over game reviews. Fine-tuned from distilbert-base-uncased on ~4.2M Steam reviews, exported to ONNX and dynamically quantised to int8 for browser inference.

Labels: negative (0), positive (1).

Results

Quantisation cost nothing here, unlike the companion game-ID model.

Both numbers are real and they measure different things. The higher one is on the corpus's natural class balance, which is heavily positive; the lower one is on a balanced sample, which is the harder and more honest test. Mean P(positive) is 0.969 for positive reviews against 0.286 for negative ones.

Size

268 MB fp32 → 67 MB int8.

Usage

import { pipeline } from "@huggingface/transformers";

const sentiment = await pipeline(
  "text-classification",
  "clifftonowen/game-review-sentiment",
  { dtype: "q8" },
);

await sentiment("refunded after twenty minutes, it stutters on everything");
// → negative, 0.985

Live demo: https://clifftonowen.me/demos

Provenance note

Exported from training checkpoint 125000, which trainer_state.json records as best_model_checkpoint (eval accuracy 0.9475, eval F1 0.9473). An earlier saved copy of this model was found to be degenerate, emitting a near-constant logit and scoring chance on balanced data, so it was discarded in favour of the checkpoint. Worth stating plainly: the training was sound, one exported artefact was not.

Limitations

Steam review register specifically. Short, sarcastic, or mixed reviews are the weak spot, which is where the balanced-sample number comes from.