game-review-game-id

Game Review → Game Identification (int8 ONNX)

Given the text of a game review, predict which of 1,232 games it is about, using the review text alone with no metadata.

Fine-tuned from distilbert-base-uncased on ~4.2M Steam reviews, restricted to the 1,232 games with at least 501 reviews. Exported to ONNX and dynamically quantised to int8 so it can run in a browser.

Results

Measured on a held-out sample from the source corpus. The 1.4-point drop is the cost of quantisation, measured rather than assumed. 60% across 1,232 classes is roughly 750× better than chance.

Errors are usually sibling titles: the model's nearest neighbours for Counter-Strike are Counter-Strike: Source and Counter-Strike: Condition Zero, and for Fallout: New Vegas they are Fallout 3 and Oblivion. It learned franchise and genre structure it was never told about.

Size

268 MB fp32 → 68 MB int8.

Usage (transformers.js, in the browser)

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

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

await classify("the gunplay is tight but the story drags in the back half", { top_k: 5 });
// → BioShock Infinite 31.4%, Spec Ops: The Line 17.2%, Call of Juarez Gunslinger 8.8%, …

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

Note on labels

The original training run saved the classifier with placeholder labels (LABEL_0 … LABEL_1231). The real mapping was recovered from the source corpus without retraining and verified by scoring the model with it (61.7% against a 0.08% baseline). The labels in config.json here are the recovered, verified ones.

Limitations

Trained on Steam review text, so it reflects that register and that era of games. It can only name games inside its 1,232-class vocabulary, and it will confidently pick the closest one for reviews of anything else.