pblrvo/Qwen3-8B-Game-semantic-IDs-v3

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

pblrvo/Qwen3-8B-Game-semantic-IDs-v3 is an 8 billion parameter Qwen3-based generative recommender model developed by pblrvo. It is fine-tuned to predict a user's next game, map game names to/from semantic IDs, and suggest similar games by reasoning over compressed semantic IDs of Steam game catalog items. This model is a full-parameter fine-tune, improving upon previous versions by using a redesigned dataset and more extensive training to enhance recommendation accuracy and semantic understanding.

Loading preview...

Overview

pblrvo/Qwen3-8B-Game-semantic-IDs-v3 is a generative recommender model based on the Qwen3-8B architecture. It uniquely processes Steam game catalog items by embedding and compressing them into short semantic IDs using an RQ-VAE. These semantic IDs form the model's vocabulary, allowing it to predict a user's next game, translate between game names and IDs, and suggest similar titles without directly handling raw item IDs or embeddings.

Key Capabilities

  • Generative Recommendation: Predicts a user's next game based on their history.
  • Semantic ID Mapping: Accurately converts game names to semantic IDs and vice-versa.
  • Similar Game Suggestion: Recommends games semantically similar to a given item.
  • Relational Understanding: Achieves 100% accuracy on a relatedness task, confirming its ability to reliably interpret the internal compositional structure of semantic IDs.
  • Improved Training: Benefits from a full-parameter fine-tune over two epochs on a redesigned dataset, addressing undertraining issues of previous versions.

Evaluation Highlights

The model shows significant improvements in grounding_name2id (96.80% Recall@10) and nl_preference (94.95% Recall@10) compared to its predecessor. While sequential and similar_item tasks show lower exact-target Recall@10 (1.0-2.0%), re-scoring for "any-played Hit@10" reveals the model retrieves genuinely relevant games well above chance (e.g., 58.00% for sequential), indicating strong underlying recommendation capabilities despite data limitations regarding chronological order.

Good for

  • Developing game recommendation systems that leverage semantic understanding.
  • Applications requiring mapping between natural language game titles and abstract identifiers.
  • Research into generative recommender models and the use of compressed semantic representations.

Finetunes
1 models