gamespace.2

gamespace.2 embeds video games into a 128-dimensional space where games that the same people enjoy sit close together. It powers the GameMerge recommender. Output vectors are L2-normalised, so a dot product is the cosine similarity.

It is all-MiniLM-L6-v2 fine-tuned on Steam store data, like gamespace.1. It also learns from player behaviour: pairs of games that the same Steam reviewers engaged with are positive examples. The model still reads only a game's store text, so it works for any game, including new ones with no players yet.

Steam next-game benchmark

The benchmark follows the protocol of SASRec (Kang & McAuley, ICDM 2018) on their Steam review crawl. Each user's most recent reviewed game is held out and ranked against 100 randomly sampled games, and also against the whole catalog. Text models score candidates by cosine similarity to the mean of the user's earlier games. Games no longer on the Steam store have no text to embed, so they were removed before the five-review filter. That leaves 250,955 users, 10,318 games and 3.05M interactions. The paper had 334,730 users and 13,047 games. Popularity and SASRec were re-run on this exact split.

Model	Learns from user histories	Hit@10, 100 negatives	NDCG@10, 100 negatives	Hit@10, all games	NDCG@10, all games
SASRec, reproduced	yes	0.836	0.581	0.115	0.060
Popularity	yes	0.688	0.420	0.050	0.025
gamespace.2, soft-max over history	pairs from training histories	0.518	0.308	0.046	0.025
gamespace.2, mean of history	pairs from training histories	0.511	0.301	0.034	0.018
gamespace.2 + CLIP header art	pairs from training histories	0.520	0.309	0.037	0.019
gamespace.1	no	0.438	0.251	0.027	0.014
e5-large-v2	no	0.404	0.234	0.030	0.016
Qwen3-Embedding-0.6B	no	0.336	0.185	0.018	0.010
bge-large-en-v1.5	no	0.334	0.191	0.026	0.014
all-MiniLM-L6-v2	no	0.306	0.162	0.017	0.010
CLIP header art only	no	0.288	0.158	0.014	0.007

For reference, the SASRec paper reports 0.873 and 0.631 on its full 13,047-game set.

The soft-max row scores each candidate by a soft maximum (log-mean-exp, temperature 0.1, chosen on the validation split) of its cosine similarity to each earlier game, instead of similarity to their mean. Mixed histories are not pulled toward one average point, which matters most when ranking the whole catalog. GameMerge uses it to combine your picks.

gamespace.2 is the best of the text embedding models, ahead of much larger general-purpose embedders. Sequential recommenders trained on the interaction data remain far stronger at predicting the next game, because they also learn popularity and order. gamespace.2 trades that for working on any game from its description alone.

Training

• Data. The Steam Games Dataset has 147,352 apps. Soundtracks, demos, playtests, servers, SDKs, and empty records were removed, leaving 138,920 games.
• Objective. Contrastive training uses CachedMultipleNegativesRankingLoss with 128 in-batch negatives. There are three kinds of positive pairs:
  • 70,000 pairs are two randomly degraded views of the same game.
  • 70,000 pairs are games whose vote-weighted user-tag profiles have a cosine of at least 0.6.
  • 66,904 pairs are co-liked games. There are 36,371 distinct pairs, each shown in both orders. They are mined from the benchmark's training histories only, so no validation or test game is used. A pair must be shared by at least 5 reviewers, it is scored by normalised PMI, and each game keeps its 10 strongest partners.
• Settings. The projection is Dense(384→128), initialised from PCA. Training runs one epoch at learning rate 3e-5, with 128-token documents.

A looser pair set, with at least 3 shared reviewers and 20 partners per game, scored lower on validation, with 0.477 against 0.503 Hit@10. Pair quality matters more than quantity.

Usage

The model reads one short document per game. Build it with docs.py from this repo, using the same format as gamespace.1:

Hollow Knight. Genres: Action, Adventure, Indie. Tags: Metroidvania, Souls-like, ... Modes: Single-player. Developer: Team Cherry. Forge your own path...

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("GulkoA/gamespace.2")
vec = model.encode("Hollow Knight. Genres: Action, Adventure, Indie. Tags: Metroidvania, Souls-like.")

In the browser

onnx/model_quantized.onnx is 22 MB, per-channel int8, with cosine 0.995 to the PyTorch model. It runs with transformers.js. Apply the 128x384 projection in dense.bin, which holds float32 weights in row-major order followed by the 128 biases, to the mean-pooled output:

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

const extractor = await pipeline("feature-extraction", "GulkoA/gamespace.2", { dtype: "q8" });
const w = new Float32Array(await (await fetch("https://huggingface.co/GulkoA/gamespace.2/resolve/main/dense.bin")).arrayBuffer());
const { data: x } = await extractor("cozy farming game with fishing", { pooling: "mean", normalize: false });
const y = Array.from({ length: 128 }, (_, i) => w[128 * 384 + i] + x.reduce((s, xj, j) => s + w[i * 384 + j] * xj, 0));
const norm = Math.hypot(...y);
const vec = y.map((v) => v / norm);

Limitations

• The behaviour pairs come from reviews written between 2010 and January 2018. Newer games benefit only through what the model generalises from text.
• Steam display names identify reviewers, and different people can share a name.
• Similarity ignores popularity and quality, so obscure games rank alongside well-known ones.

License

The model is licensed under Apache 2.0, the same as the base model. The store data is MIT-licensed. The review crawl is from the UCSD Steam datasets (Kang & McAuley, 2018).