Downloads · 30 days
0
siglip2-bilingual-ddp-lora
ShivRamSaud/siglip2-bilingual-ddp-lora
siglip2-bilingual-ddp-lora is a machine learning model from ShivRamSaud. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Fine-tuned from google/siglip2-so400m-patch14-384 with LoRA r=8 alpha=16. In-notebook training (no torchrun, Papermill-debuggable). HF-only: no saves to /kaggle/working/outputs except temp files for HF upload, then de…
Access
Public
Updated Sep 23, 2026
Repo size
50.2 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt9.1 GB · 99%
At a glance
- Access
- Public
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026
- SHA
- e4e1f661
Model card
From the Hugging Face model README
SigLIP2 DDP LoRA Mean/Mean (EN-Optimized) -- IN-NOTEBOOK HF-ONLY
Fine-tuned from google/siglip2-so400m-patch14-384 with LoRA r=8 alpha=16. In-notebook training (no torchrun, Papermill-debuggable). HF-only: no saves to /kaggle/working/outputs except temp files for HF upload, then delete local. All artifacts pushed to HF repo ShivRamSaud/siglip2-bilingual-ddp-lora.
- Best adapter:
checkpoints/best_adapter/(HF) -> localcheckpoints/siglip2_lora_best_adapter/deleted after upload - Metrics:
siglip2_ddp_lora_results.csv+final_eval_best_adapter.csv(HF, deleted local) - Frozen baseline was 27.68% T2V EN R@1 (VAL 638/112 val_md5=2a372b4d28c5, mean/mean, 4 frames).
- LoRA adapters are HF
save_pretrainedformat -- load via PeftModel.from_pretrained(HF_REPO_ID, subfolder="checkpoints/best_adapter") - Nepali trained via bilingual alternation (EN/NE per step), all metrics R@1/5/10 MedR MRR mAP nDCG@10 DCG@10 MAP@10 HITS@1 Margin covered.
- KAGGLE paths only: KAGGLE_CAPTIONS=/kaggle/input/datasets/shivramsaud/final-retrival/outputs, KAGGLE_VIDEOS=/kaggle/input/datasets/shivramsaud/video-clip/annotated, Kaggle-only
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026