Charformer: Fast Character Transformers via Gradient-based Subword Tokenization
Paper • 2106.12672 • Published
Charformer Turkish Base Model
Still in development phase!!! Weights are updating regularly
Character-level encoder-decoder T5 transformer.
Model arhitecture based on https://arxiv.org/abs/2106.12672
Parameters
- max_subword_block_width = 4
- downsample_rate = 2
- d_model = 768
- n_head = 12
- encoder_num_layers = 12
- decoder_num_layers = 12
- window = 1024
Usage
This model is not tokenizer-based.
Use vocab.json for char → id mapping.
Author
This model was developed by Orkun Gedik with the academic and technical support of Gazi University AI center and computer engineering department, Ankara. The project benefited from Gazi University’s intellectual contributions which played an important role in the design, training, and evaluation of the model.
- Downloads last month
- 52
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support