heimai/Industrial_and_Scientific_Qwen2.5_0.5B_sft
heimai/Industrial_and_Scientific_Qwen2.5_0.5B_sft is a 0.5 billion parameter generative recommendation model, MiniOneRec, built upon the Qwen2.5 base architecture. It utilizes Semantic ID (RQ-VAE) technology to encode items and transforms recommendation tasks into sequence generation. The model is fine-tuned with SFT for instruction adherence and optimized with GRPO reinforcement learning for ranking quality, achieving a Hit Rate of 17.27% and NDCG of 8.18% at @50.
Loading preview...
MiniOneRec: Generative Recommendation Model
MiniOneRec is a generative recommendation system model based on the Qwen2.5-0.5B architecture. It innovatively converts recommendation tasks into sequence generation by encoding items using Semantic ID (RQ-VAE) technology. This approach allows the model to predict item IDs as a generative task.
Key Capabilities
- Generative Recommendation: Predicts item IDs by generating sequences, moving beyond traditional collaborative filtering.
- Semantic ID Encoding: Utilizes RQ-VAE to create semantic IDs for items, capturing their underlying meaning and relationships.
- Instruction Fine-tuning (SFT): Trained to understand and follow recommendation instructions, ensuring structured output.
- Reinforcement Learning Optimization (GRPO): Employs GRPO with ranking rewards to enhance the quality of recommendations and improve item ordering.
Performance Metrics
Evaluated using Beam Search (num_beams=50), MiniOneRec demonstrates competitive performance for its size:
- Hit Rate (HR): 17.27% at @50
- NDCG (Ranking Quality): 8.18% at @50
Good for
- Developing generative recommendation systems where item IDs are predicted as sequences.
- Applications requiring semantic understanding of items through encoded IDs.
- Scenarios benefiting from a small, efficient model (0.5B parameters) for recommendation tasks.
- Research and development in reinforcement learning for recommendation.