Downloads · 30 days
0
kuairand-semantic-id-rqvae
plumliu/kuairand-semantic-id-rqvae
kuairand-semantic-id-rqvae is a machine learning model from plumliu. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as apache-2.0.
This repository hosts the lightweight RQ-VAE checkpoint used to discretize KuaiRand-27K item hybrid embeddings into three-level Semantic IDs.
Access
Public
Updated Jun 22, 2026
Repo size
2.9 MB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.pt2.9 MB · 99%
At a glance
- Library
- pytorch
- License
- apache-2.0
- Access
- Public
- Created
- Jun 22, 2026
- Updated
- Jun 22, 2026
- SHA
- 6d22e755
Model card
From the Hugging Face model README
KuaiRand Semantic ID RQ-VAE
This repository hosts the lightweight RQ-VAE checkpoint used to discretize KuaiRand-27K item hybrid embeddings into three-level Semantic IDs.
It is part of a small public release bundle:
- GitHub landing repository:
plumliu/kuai-recexplain-lora - LoRA adapter:
plumliu/qwen35-9b-kuairand-recexplain-lora - SFT/DPO and Semantic ID dataset:
plumliu/kuairand-recexplain-sft-dpo-data - RQ-VAE checkpoint: this repository
What This Model Does
The model maps a 256-dimensional hybrid item embedding into a three-token Semantic ID:
hybrid_item_embedding[256]
-> RQ-VAE residual quantization
-> <SID_0_x> <SID_1_y> <SID_2_z>
The hybrid embedding was built from standardized item text and user-behavior signals. Semantic IDs are used as compact structured item tokens for recommendation explanation post-training.
Released Files
| Path | Description |
|---|---|
rqvae_best.pt | Best RQ-VAE checkpoint for the formal KuaiRand-27K item universe. |
config/formal_v100.yaml | Compact pipeline config used for the Semantic ID stage. |
stats/rqvae_semantic_id_summary.json | Training and reconstruction summary. |
stats/rqvae_semantic_id_summary.md | Human-readable training summary. |
stats/semantic_id_health_check.json | Semantic ID neighborhood and collision health check. |
stats/semantic_id_health_check.md | Human-readable health check summary. |
The generated Semantic ID mapping and the source hybrid embeddings are released in the dataset repository, not here:
plumliu/kuairand-recexplain-sft-dpo-data
Architecture Summary
- Input dimension:
256 - Number of residual codebooks:
3 - Codebook size per level:
256 - Output token pattern:
<SID_0_*> <SID_1_*> <SID_2_*>
Formal validation results:
- Best validation reconstruction MSE:
0.00191175 - Codebook usage: level 0
256/256, level 1256/256, level 2255/256 - Full three-level unique Semantic IDs:
218,282 - SID collisions:
3,325items, max bucket size5
Neighborhood checks showed same-prefix pairs were much closer than random pairs:
| Check | Same-prefix | Random |
|---|---|---|
| Level-1 prefix hybrid cosine | 0.417966 | 0.239881 |
| Level-2 prefix hybrid cosine | 0.493802 | 0.239881 |
| Level-1 prefix category match | 0.311200 | 0.061120 |
Intended Use
This checkpoint is intended for research reproduction and inspection of the Semantic ID construction stage. It is not a standalone recommender model and does not include KuaiRand raw logs, text embeddings, behavior embeddings, or item metadata.
Related Artifacts
- Use the dataset repo for
semantic_id_mapping.parquet,semantic_id_levels.int16.npy, hybrid embeddings, and rewritten user sequences. - Use the LoRA adapter repo for the final recommendation explanation model.
- Use the GitHub landing repo for compact method notes, examples, and evaluation summaries.
- Library
- pytorch
- License
- apache-2.0
- Languages
- zh, en
- Created
- Jun 22, 2026
- Updated
- Jun 22, 2026