Support distilroberta-base in spacy-curated-transformers

[znadrich-qf]

spacy-curated-transformers supports roberta-base but not distilroberta-base from Huggingface - could support for distilroberta-base be added? I understand that spacy-transformers supports distilroberta-base, but we'd like to use the same span getters from spacy-curated-transformers which seem incompatible from the two libraries.

Specifically it would seem that the span getters from spacy-curated-transformers operate on sub-tokens while the span getters from spacy-transformers operate on spaCy tokens, meaning that we have less control in specifying how long our context window is.

[znadrich-qf]

Actually this is already supported if you do

[components.transformer.model]
@architectures = "spacy-curated-transformers.RobertaTransformer.v1"
vocab_size = 50265
hidden_width = 768
piece_encoder = {"@architectures":"spacy-curated-transformers.ByteBpeEncoder.v1"}
attention_probs_dropout_prob = 0.1
hidden_act = "gelu"
hidden_dropout_prob = 0.1
intermediate_width = 3072
layer_norm_eps = 0.00001
max_position_embeddings = 514
model_max_length = 512
num_attention_heads = 12
num_hidden_layers = 6
padding_idx = 1
type_vocab_size = 1
torchscript = false
mixed_precision = true
wrapped_listener = null

[initialize.components.transformer]

[initialize.components.transformer.encoder_loader]
@model_loaders = "spacy-curated-transformers.HFTransformerEncoderLoader.v1"
name = "distilroberta-base"
revision = "main"

[initialize.components.transformer.piecer_loader]
@model_loaders = "spacy-curated-transformers.HFPieceEncoderLoader.v1"
name = "distilroberta-base"
revision = "main"