LLM-jp CLIP ViT Large/14
LLM-jp CLIP Large/14 combines a Japanese RoBERTa text encoder with ViT-L/14. It was trained through OpenCLIP on translated ReLAION image-text pairs.
Research release
The captions were translated with Gemma
Gemma-2-9B-it translated the source English captions into Japanese. The authors state that this translation use does not make the resulting CLIP model a Gemma derivative under their reading of the terms. That is the developers' stated interpretation.LLM-jp CLIP ViT Large/14 official model card
Use embedding comparisons for retrieval
Practical guidance. Compare this larger image encoder with Base/16 on the same image and text pairs. Its parameter increase is in a different retrieval architecture from a generative vision-language model.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | llm-jp-clip-vit-large-patch14 |
| Parameters | 467 million.LLM-jp CLIP ViT Large/14 official model card |
| Training data | 1.45 billion downloaded image-text pairs, processed for nine epochs.LLM-jp CLIP ViT Large/14 official model card |
| License | Apache 2.0.LLM-jp CLIP ViT Large/14 official model card |
Related models
- LLM-jp CLIP ViT Base/16
LLM-jp / NII LLMC · LLM-jp CLIP
A 248 million-parameter Japanese image-text encoder for retrieval and zero-shot image classification.
Launch date not confirmedResearch releaseDownloadable A 375M-parameter SigLIP 2 derivative tuned on Japanese WAON image-text pairs.
Launch date not confirmedResearch releaseDownloadable