The aim is to train model for NER-extraction with score. Spacy documentation recommends to use span_categorizer in order to get score values. I’ve configured spacy pipeline which consists of two components: transformer and SpanCategorizer. It might be singlelabel mode because of data specific. So base part of the config is :

[nlp]
...
pipeline = ["transformer","spancat_singlelabel"]
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}
vectors = {"@vectors":"spacy.Vectors.v1"}

[components]

[components.spancat_singlelabel]
factory = "spancat_singlelabel"
allow_overlap = true
negative_weight = 0.6
scorer = {"@scorers":"spacy.spancat_scorer.v1"}
spans_key = "sc"

[components.spancat_singlelabel.model]
@architectures = "spacy.SpanCategorizer.v1"

[components.spancat_singlelabel.model.reducer]
@layers = "spacy.mean_max_reducer.v1"
hidden_size = 128

[components.spancat_singlelabel.model.scorer]
@layers = "Softmax.v2"
nO = null
nI = null
init_W = null
init_b = null

[components.spancat_singlelabel.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "*"

[components.spancat_singlelabel.suggester]
@misc = "spacy.ngram_suggester.v1"
sizes = [1,2,3,4,5,6,7]

[components.transformer]
factory = "transformer"
max_batch_items = 4096
set_extra_annotations = {"@annotation_setters":"spacy-transformers.null_annotation_setter.v1"}

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "Geotrend/distilbert-base-25lang-cased"
mixed_precision = false

[components.transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 200
stride = 10

But in this case transformer each epoch gets zero losses. If I change scorer layer (section components.spancat_singlelabel.model.scorer) to spacy.LinearLogistic.v1 it goes ok.

[nlp]
...
pipeline = ["transformer","spancat_singlelabel"]
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}
vectors = {"@vectors":"spacy.Vectors.v1"}

[components]

[components.spancat_singlelabel]
factory = "spancat_singlelabel"
allow_overlap = true
negative_weight = 0.6
scorer = {"@scorers":"spacy.spancat_scorer.v1"}
spans_key = "sc"

[components.spancat_singlelabel.model]
@architectures = "spacy.SpanCategorizer.v1"

[components.spancat_singlelabel.model.reducer]
@layers = "spacy.mean_max_reducer.v1"
hidden_size = 128

[components.spancat_singlelabel.model.scorer]
@layers = "spacy.LinearLogistic.v1"
nO = null
nI = null

[components.spancat_singlelabel.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "*"

[components.spancat_singlelabel.suggester]
@misc = "spacy.ngram_suggester.v1"
sizes = [1,2,3,4,5,6,7]

[components.transformer]
factory = "transformer"
max_batch_items = 4096
set_extra_annotations = {"@annotation_setters":"spacy-transformers.null_annotation_setter.v1"}

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "Geotrend/distilbert-base-25lang-cased"
mixed_precision = false

[components.transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 200
stride = 10

Is it possible to use SoftMax layer for SpanCategorizer in single_mode with transformer in pipeline?