Tawhidhere/OneRec-Think
🤖 Enhance recommendations with OneRec-Think, a framework that combines dialogue, reasoning, and personalization for precise and engaging user experiences.
What's novel
🤖 Enhance recommendations with OneRec-Think, a framework that combines dialogue, reasoning, and personalization for precise and engaging user experiences.
Code Analysis
11 files read · 3 roundsA 3-stage LLM-based sequential recommendation pipeline on the Beauty dataset that expands Qwen3-1.7B's vocabulary with semantic ID tokens, aligns them via embedding training, fine-tunes with LoRA for next-item prediction, and adds chain-of-thought reasoning with trie-constrained beam search decoding
Strengths
The trie-constrained decoding for guaranteed valid SID generation and the iterative data reconstruction loop in the RA stage show genuine research thinking. The pipeline is complete end-to-end from vocab expansion through multi-GPU parallel evaluation.
Weaknesses
Severe code duplication across all three training scripts (CustomDataCollator, token utilities, dataset prep are copy-pasted), zero unit tests, and the README is entirely fabricated marketing copy that describes a non-existent consumer GUI app. The CustomDataCollator's decode-then-rescan approach for loss masking is fragile and inefficient.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
7
Contributors
2
Files
29
Active weeks
3
Repository
Language
Python
Stars
1
Forks
0
License
Apache-2.0