Remove stage-specific wording from model card

README.md

@@ -1,52 +1,52 @@
 ---
-license: mit
-language: en
 tags:
-  - recommendation
-  - rq-vae
-  - stage-2
-  - diger
-  - arxiv:2601.19711
+- diger
+- rq-vae
+- generative-recommendation
+- semantic-id
+- recommendation
+pipeline_tag: feature-extraction
 ---
 
-# DIGER RQ-VAE Stage-2 Checkpoint (instruments)
+# DIGER RQ-VAE Checkpoint (instruments)
 
-This repository provides the Stage-2 RQ-VAE checkpoint for **instruments** used in the paper:
-*Fu et al., "Differentiable Semantic ID for Generative Recommendation" (arXiv:2601.19711, 2026)*.
+This repository provides the pretrained RQ-VAE checkpoint for **instruments** used in the paper:
+
+**DIGER: Differentiable Semantic IDs for Generative Recommendation**
 
 ## Files
 
-- `best_collision_model.pth`: Stage-2 checkpoint from the DIGER RQ-VAE pretraining pipeline.
+- `best_collision_model.pth`: released checkpoint from the DIGER RQ-VAE pretraining pipeline.
+
+## Usage
 
-## Checkpoint settings
+Download with `huggingface_hub`:
 
-- epochs: 10000
-- batch_size: 2048
-- lr: 1e-3
-- weight_decay: 1e-4
-- e_dim: 256
-- num_emb_list: [256, 256, 256]
-- layers: [2048, 1024, 512]
-- beta: 0.25
-- sk_epsilons: [0.003, 0.003, 0.003]
-- sk_iters: 50
-- vq_type: vq
-- loss_type: mse
-- dist: l2
+```python
+from huggingface_hub import hf_hub_download
 
-## Code
+ckpt_path = hf_hub_download(
+    repo_id="junchenfu/diger-rqvae-instruments",
+    filename="best_collision_model.pth",
+)
+```
 
-Code and scripts are in the DIGER repository:
+The checkpoint can be loaded with the RQ-VAE implementation in the DIGER GitHub repository:
 
-- https://github.com/junchen-fu/DIGER
+https://github.com/junchen-fu/DIGER
 
-Embedding inputs used for this stage follow the procedure described in:
-- https://github.com/honghuibao2000/letter
+Please refer to the repository README for the full training configuration and commands.
 
-## How to load
+## Embeddings
 
-```python
-import torch
-ckpt = torch.load('best_collision_model.pth', map_location='cpu')
-print(ckpt.keys())
-```
+LLaMA embeddings used by DIGER should be generated following the procedure described in:
+
+https://github.com/honghuibao2000/letter
+
+## Dataset Note
+
+The underlying recommendation datasets are publicly available from their original sources. Processed interaction data are not hosted in this model repository; they can be regenerated following the DIGER codebase.
+
+## Citation
+
+If you use this checkpoint, please cite the DIGER paper.