# Understanding config file output by textcat based on ja\_core\_news\_trf

**URL:** <https://support.prodi.gy/t/understanding-config-file-output-by-textcat-based-on-ja-core-news-trf/7625>\
**Category:** Uncategorized\
**Tags:** textcat\
**Created:** [October 22, 2025, 7:16am UTC](https://support.prodi.gy/t/understanding-config-file-output-by-textcat-based-on-ja-core-news-trf/7625 "2025-10-22T07:16:24Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![jhandsel](https://sea2.discourse-cdn.com/flex020/user_avatar/support.prodi.gy/jhandsel/32/4621_2.png) [@jhandsel](https://support.prodi.gy/u/jhandsel)\
**Post date:** [October 22, 2025, 7:16am UTC](https://support.prodi.gy/t/understanding-config-file-output-by-textcat-based-on-ja-core-news-trf/7625/1 "2025-10-22T07:16:24Z")

</div>

I’m trying to understand the config file created by Prodigy when training textcat on top of the Japanese transformer model. I’m training with the following command:

```bash
prodigy train my_model --textcat-multilabel my_dataset --base-model ja_core_news_trf

```

Looking at the config output from the above command, I see a textcat component, but I don’t see any listeners that would take features from the transformer. **Am I correct in thinking that this config results in a simple bag of words classifier that doesn't make use of the transformer layer?**

The full config file:

```python
[paths]
train = null
dev = null
vectors = null
init_tok2vec = null

[system]
gpu_allocator = null
seed = 0

[nlp]
lang = "ja"
pipeline = ["transformer","morphologizer","parser","attribute_ruler","ner","textcat_multilabel"]
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
batch_size = 64
vectors = {"@vectors":"spacy.Vectors.v1"}

[nlp.tokenizer]
@tokenizers = "spacy.ja.JapaneseTokenizer"
split_mode = null

[components]

[components.attribute_ruler]
source = "ja_core_news_trf"

[components.morphologizer]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.ner]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.parser]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.textcat_multilabel]
factory = "textcat_multilabel"
scorer = {"@scorers":"spacy.textcat_multilabel_scorer.v2"}
threshold = 0.5

[components.textcat_multilabel.model]
@architectures = "spacy.TextCatBOW.v3"
exclusive_classes = false
length = 262144
ngram_size = 1
no_output_layer = false
nO = null

[components.transformer]
source = "ja_core_news_trf"

[corpora]

[corpora.dev]
@readers = "spacy.Corpus.v1"
path = ${paths.dev}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[training]
train_corpus = "corpora.train"
dev_corpus = "corpora.dev"
seed = ${system:seed}
gpu_allocator = ${system:gpu_allocator}
dropout = 0.1
accumulate_gradient = 3
patience = 5000
max_epochs = 0
max_steps = 20000
eval_frequency = 1000
frozen_components = ["morphologizer","parser","attribute_ruler","ner"]
before_to_disk = null
annotating_components = []
before_update = null

[training.batcher]
@batchers = "spacy.batch_by_words.v1"
discard_oversize = false
size = 2000
tolerance = 0.2
get_length = null

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = true
eps = 0.00000001

[training.optimizer.learn_rate]
@schedules = "warmup_linear.v1"
warmup_steps = 250
total_steps = 20000
initial_rate = 0.00005

[training.score_weights]
pos_acc = null
morph_micro_f = null
morph_per_feat = null
dep_uas = null
dep_las = null
dep_las_per_type = null
sents_p = null
sents_r = null
sents_f = null
ents_f = null
ents_p = null
ents_r = null
ents_per_type = null
morph_acc = 0.11
speed = 0.0

[pretraining]

[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
after_init = null

[initialize.before_init]
@callbacks = "spacy.copy_from_base_model.v1"
tokenizer = "ja_core_news_trf"
vocab = "ja_core_news_trf"

[initialize.components]

[initialize.components.morphologizer]

[initialize.components.morphologizer.labels]
@readers = "spacy.read_labels.v1"
path = "negative_teach_biased_spacy/labels/morphologizer.json"

[initialize.components.ner]

[initialize.components.ner.labels]
@readers = "spacy.read_labels.v1"
path = "negative_teach_biased_spacy/labels/ner.json"

[initialize.components.parser]

[initialize.components.parser.labels]
@readers = "spacy.read_labels.v1"
path = "negative_teach_biased_spacy/labels/parser.json"

[initialize.components.textcat_multilabel]

[initialize.components.textcat_multilabel.labels]
@readers = "spacy.read_labels.v1"
path = "negative_teach_biased_spacy/labels/textcat_multilabel.json"

[initialize.tokenizer]

```

---

<div class="post-metadata">

**Author:** ![magdaaniol](https://sea2.discourse-cdn.com/flex020/user_avatar/support.prodi.gy/magdaaniol/32/2787_2.png) [@magdaaniol](https://support.prodi.gy/u/magdaaniol)\
**Post date:** [November 4, 2025, 10:46am UTC](https://support.prodi.gy/t/understanding-config-file-output-by-textcat-based-on-ja-core-news-trf/7625/2 "2025-11-04T10:46:10Z")

</div>

Hi @jhandsel,

You’re completely right — the reason your config ends up with a simple bag-of-words (`TextCatBOW.v3`) textcat component instead of a transformer-aware one is because of how Prodigy decides whether to optimize for efficiency or accuracy.  
Japanese transformer models like `ja_core_news_trf` don’t include pretrained word vectors, so Prodigy mistakenly generates the “efficiency” setup (the older TextCatBOW model) instead of the “accuracy” setup that connects to the transformer via embedding layer using e.g [spacy.TextCatEnsemble.v2](https://spacy.io/api/architectures#TextCatEnsemble) architecture.

The code responsible for this was originally written for older pipelines (e.g. English `en_core_web_lg`) that bundled pretrained word vectors, but it obviously is not correct for the newer Japanese transformer models.

Until the Prodigy logic is updated, you have two options:

1. Pass the corrected config manually to Prodigy `train` recipe:

```python
prodigy train my_model --textcat-multilabel my_dataset --base-model ja_core_news_trf --config correct_transformer_config.cfg

```

1. Export your annotations from Prodigy to spaCy DocBin format using [`data-to-spacy`](https://prodi.gy/docs/recipes#data-to-spacy) command and train directly with spaCy using spaCy [`train`](https://spacy.io/api/cli#train) command.

An example of the correct config for training **with Prodigy** using `spacy.TextCatEnsemble.v2` architecture that listens to the shared transformer embedding layer would be:

```python
[paths]
train = null
dev = null
vectors = null
init_tok2vec = null

[system]
gpu_allocator = "pytorch"
seed = 0

[nlp]
lang = "ja"
pipeline = ["transformer","morphologizer","parser","attribute_ruler","ner","textcat_multilabel"]
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
batch_size = 128

[nlp.tokenizer]
@tokenizers = "spacy.ja.JapaneseTokenizer"
split_mode = null

[components]

[components.attribute_ruler]
source = "ja_core_news_trf"

[components.morphologizer]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.ner]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.parser]
source = "ja_core_news_trf"
replace_listeners = ["model.tok2vec"]

[components.textcat_multilabel]
factory = "textcat_multilabel"
scorer = {"@scorers":"spacy.textcat_multilabel_scorer.v2"}
threshold = 0.5

[components.textcat_multilabel.model]
@architectures = "spacy.TextCatEnsemble.v2"
nO = null

[components.textcat_multilabel.model.linear_model]
@architectures = "spacy.TextCatBOW.v3"
exclusive_classes = false
length = 262144
ngram_size = 1
no_output_layer = false
nO = null

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

[components.transformer]
source = "ja_core_news_trf"

[corpora]
@readers = "prodigy.MergedCorpus.v1"
eval_split = 0.2
sample_size = 1.0

[corpora.textcat_multilabel]
@readers = "prodigy.TextCatCorpus.v1"
datasets = ["my_dataset"]
eval_datasets = []
exclusive = false

[training]
train_corpus = "corpora.train"
dev_corpus = "corpora.dev"
seed = ${system:seed}
gpu_allocator = ${system:gpu_allocator}
dropout = 0.1
accumulate_gradient = 3
patience = 5000
max_epochs = 0
max_steps = 20000
eval_frequency = 1000
frozen_components = ["morphologizer","parser","attribute_ruler","ner"]
before_to_disk = null
annotating_components = []
before_update = null

[training.batcher]
@batchers = "spacy.batch_by_words.v1"
discard_oversize = false
size = 2000
tolerance = 0.2
get_length = null

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = true
eps = 0.00000001

[training.optimizer.learn_rate]
@schedules = "warmup_linear.v1"
warmup_steps = 250
total_steps = 20000
initial_rate = 0.00005

[training.score_weights]
pos_acc = null
morph_micro_f = null
morph_per_feat = null
dep_uas = null
dep_las = null
dep_las_per_type = null
sents_p = null
sents_r = null
sents_f = null
ents_f = null
ents_p = null
ents_r = null
ents_per_type = null
morph_acc = 0.11
speed = 0.0

[pretraining]

[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
after_init = null

[initialize.before_init]
@callbacks = "spacy.copy_from_base_model.v1"
tokenizer = "ja_core_news_trf"
vocab = "ja_core_news_trf"

[initialize.components]

[initialize.tokenizer]

```

---

<div class="post-metadata">

**Author:** ![jhandsel](https://sea2.discourse-cdn.com/flex020/user_avatar/support.prodi.gy/jhandsel/32/4621_2.png) [@jhandsel](https://support.prodi.gy/u/jhandsel)\
**Post date:** [November 17, 2025, 2:57am UTC](https://support.prodi.gy/t/understanding-config-file-output-by-textcat-based-on-ja-core-news-trf/7625/3 "2025-11-17T02:57:23Z")

</div>

Thanks for confirming @magdaaniol

I have been working around the issue with a custom config file.