Cosmos ColBERT

Meet the smallest BERT for Cosmos; yet, it's a functional text encoder for booru tags.

How much can we scale it down?

In float16, the model size is almost 69MB. The model supports up to 512 tokens.

Natural language instruction

ColBERT uses a simple position embedding rather than RoPE. It is only expected to work with comma-separated tags.

Inference

bert = CosmosColBert.from_pretrained('nightknocker/cosmos-colbert-34m')
tokenizer = AutoTokenizer.from_pretrained('nightknocker/cosmos-colbert-34m')
inputs = tokenizer(text, return_tensors='pt').to('cuda')
crossattn_emb = bert.forward(**inputs, return_dict=True).last_hidden_state

Datasets

  • anime-art-multicaptions (multicharacter interactions)
  • art movement, dress/hair/skin color, fashion, hairdo tags
  • eyes
Downloads last month
23
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for nightknocker/cosmos-colbert-34m

Finetuned
(6)
this model