This PR adds the initial models for calamanCy. For the first few versions, I think it's better to just have a _md to _trf model. There are also some design decisions that I used when training the pipelines. Here are the notable ones:
I merged the two UD treebanks because each one is too small to train a proper model from.
For the large model, I used the fastText vectors (714k keys). It was trained on CommonCrawl and should be useful for most applications.
For the transformer model, I am using jcblaise/roberta-tagalog-base. During my experiments, there's marginal difference between the base and large RoBERTa models, so I opted for the smaller version. This resulted to a slightly "leaner" pipeline (700 MB).
Please note that the Ugnayan dataset has a Non-commercial license. The TLUnified dataset is also GNU GPL, so it might be best to mention this when sharing to folks. I'd definitely want to rant about licensing, but that would be a topic for another blog post.
Models and pipelines
Model
Pipelines
Description
tl_calamancy_md (73.7 MB)
tok2vec, tagger, morphologizer, parser, ner
CPU-optimized Tagalog NLP model. Pretrained using the TLUnified dataset. Using floret vectors (50k keys)
tl_calamancy_lg (431.9 MB)
tok2vec, tagger, morphologizer, parser, ner
CPU-optimized large Tagalog NLP model. Pretrained using the TLUnified dataset. Using fastText vectors (714k)
tl_calamancy_trf (775.6 MB)
transformer, tagger, parser, ner
GPU-optimized transformer Tagalog NLP model. Uses roberta-tagalog-base as context vectors.
Description
Closes #16
This PR adds the initial models for calamanCy. For the first few versions, I think it's better to just have a
_md
to_trf
model. There are also some design decisions that I used when training the pipelines. Here are the notable ones:jcblaise/roberta-tagalog-base
. During my experiments, there's marginal difference between thebase
andlarge
RoBERTa models, so I opted for the smaller version. This resulted to a slightly "leaner" pipeline (700 MB).Models and pipelines
Data sources