Update README.md

README.md

@@ -3,8 +3,15 @@ license: mit
 base_model:
 - facebook/dinov2-small
 tags:
-- VAE,
 - SD
+- representation-learning,
+- repa
+- dinov2
+- VAE
+- stable-diffusion
+- pytroch
+datasets:
+- kingsidharth/zangei-dit-stage-1-250k
 ---
 # SD/E2E-SD VAE to DINOv2 Bridge
 GET Dino v2 features from SD VAE latents
@@ -157,7 +164,7 @@ pred_norm = F.normalize(pred_dino_features[0].float(), dim=-1)
 target_norm = F.normalize(sample_target.to(device).float(), dim=-1)
 sim = F.cosine_similarity(pred_norm, target_norm, dim=-1).mean().item()
 print(f"Average Cosine Similarity for this sample: {sim:.4f}")
-``
+```
 
 
 
@@ -179,4 +186,4 @@ print(f"Average Cosine Similarity for this sample: {sim:.4f}")
 | Best logged Retrieval@5 | `0.996364` |
 | Best logged Retrieval@10 | `0.998636` |
 
-The training log shows validation improving consistently from epoch 1 to epoch 6: cosine rose from `0.624136` to `0.759174`, NMSE fell from `0.610722` to `0.420825`, and Retrieval@1 rose from `0.751818` to `0.988182`. The run was interrupted during the final `latest.pt` save after `epoch_006.pt` had already been saved, so `best.pt` / `epoch_006.pt` are the safer checkpoints. :contentReference[oaicite:0]{index=0}
+The training log shows validation improving consistently from epoch 1 to epoch 6: cosine rose from `0.624136` to `0.759174`, NMSE fell from `0.610722` to `0.420825`, and Retrieval@1 rose from `0.751818` to `0.988182`. The run was interrupted during the final `latest.pt` save after `epoch_006.pt` had already been saved, so `best.pt` / `epoch_006.pt` are the safer checkpoints. :contentReference[oaicite:0]{index=0}