spaCy spancat won’t learn (zero F-score) while NER on same data scores 0.40 — Prodigy-generated KPI/target corpus

[marsy-41]

I am traing to train a spaCy v 3.8.7 spancat model on ~100 sustainability reports (annotated with Prodigy) to extract KPIs and targets.

An NER pipeline trained on the same data reaches F≈0.40, but spancat stays at 0.00 no matter what I try:

I ran the same code using a GPU Nvidia V100 (32GB) node, so I think that memory issues should not be the problem. I also tried spancat Transformer, with the same result.

The data seems also fine when inspecting via spacy debug:

============================ Data file validation ============================
✔ Pipeline can be initialized with data
✔ Corpus is loadable

=============================== Training stats ===============================
Language: en
Training pipeline: tok2vec, spancat
1604 training docs
401 evaluation docs
✔ No overlap between training and evaluation data
⚠ Low number of examples to train a new pipeline (1604)

============================== Vocab & Vectors ==============================
ℹ 144646 total word(s) in the data (11307 unique)
ℹ No word vectors present in the package

============================ Span Categorization ============================

Spans Key Labels

---

sc {'KPI', 'Target'}

ℹ Span characteristics for spans_key 'sc'
ℹ SD = Span Distinctiveness, BD = Boundary Distinctiveness

Span Type Length SD BD N

---

Target 11.89 0.77 1.46 447
KPI 15.80 0.60 1.29 885

---

Wgt. Average 14.48 0.66 1.35 -

ℹ Over 90% of spans have lengths of 1 -- 34 (min=1, max=162). The most
common span lengths are: 2 (1.8%), 3 (2.1%), 4 (2.48%), 5 (2.1%), 6 (2.7%), 7
(3.6%), 8 (3.38%), 9 (4.13%), 10 (5.33%), 11 (5.03%), 12 (3.53%), 13 (5.18%), 14
(3.83%), 15 (5.11%), 16 (4.43%), 17 (3.98%), 18 (4.5%), 19 (3.98%), 20 (3.15%),
21 (3.45%), 22 (2.25%), 23 (2.4%), 24 (2.48%), 25 (1.95%), 26 (1.35%), 29
(2.03%), 30 (2.1%), 31 (0.98%), 34 (0.98%). If you are using the n-gram
suggester, note that omitting infrequent n-gram lengths can greatly improve
speed and memory usage.
⚠ Spans may not be distinct from the rest of the corpus
✔ Boundary tokens are distinct from the rest of the corpus
✔ Good amount of examples for all labels
✔ Examples without occurrences available for all labels

================================== Summary ==================================
✔ 6 checks passed
⚠ 2 warnings

Here is the code I use, as well as the config file. I did play around with various span sizes to (1) avoid memory overload and (2) fit our actual span sizes:

Python 3.12.11

1. create_docbin.py

sample_docbin.py ── Python ≥3.8 · spaCy ≥3.7

import json, random, spacy
from pathlib import Path
from spacy.tokens import DocBin

IN = Path("/Users/mgulenko/Dropbox/Research/KPIs/Python/spacy_spancat/annotations.jsonl") # JSONL with {"text", "spans"}
OUT = Path("/Users/mgulenko/Dropbox/Research/KPIs/Python/spacy_spancat/sample_training_data_span") # target dir
OUT.mkdir(parents=True, exist_ok=True)

SAMPLE, DEV_RATIO = 300, 0.20 # ≤300 docs → 80/20 split
random.seed(42)

nlp = spacy.blank("en") # tokenizer only
docs = []

for rec in map(json.loads, IN.read_text().splitlines()):
doc = nlp.make_doc(rec["text"])
spans = [
doc.char_span(s["start"], s["end"], s["label"], alignment_mode="contract")
for s in rec["spans"]
]
if None not in spans: # keep only token-aligned spans
doc.spans["sc"] = spans # one span group called "sc"
docs.append(doc)

if not docs:
raise SystemExit("No valid examples")

sample = random.sample(docs, min(SAMPLE, len(docs)))
split = int(len(sample) * (1 - DEV_RATIO))

DocBin(docs=sample[:split]).to_disk(OUT / "train.spacy")
DocBin(docs=sample[split:]).to_disk(OUT / "dev.spacy")

print(f"{split} train · {len(sample)-split} dev docs → {OUT}")

2. config.cfg

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

[system]
gpu_allocator = null
seed = 0

[nlp]
lang = "en"
pipeline = ["tok2vec","spancat"]
batch_size = 5
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@Tokenizers":"spacy.Tokenizer.v1"}
vectors = {"@vectors":"spacy.Vectors.v1"}

[components]

[components.spancat]
factory = "spancat"
max_positive = null
scorer = {"@scorers":"spacy.spancat_scorer.v1"}
spans_key = "sc"
threshold = 0.5

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

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

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

[components.spancat.model.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"
width = ${components.tok2vec.model.encode.width}
upstream = "*"

[components.spancat.suggester]
@misc = "spacy.ngram_range_suggester.v1"
min_size=2
max_size=12

[components.tok2vec]
factory = "tok2vec"

[components.tok2vec.model]
@architectures = "spacy.Tok2Vec.v2"

[components.tok2vec.model.embed]
@architectures = "spacy.MultiHashEmbed.v2"
width = ${components.tok2vec.model.encode.width}
attrs = ["NORM","PREFIX","SUFFIX","SHAPE"]
rows = [5000,1000,2500,2500]
include_static_vectors = false

[components.tok2vec.model.encode]
@architectures = "spacy.MaxoutWindowEncoder.v2"
width = 96
depth = 4
window_size = 1
maxout_pieces = 3

[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]
dev_corpus = "corpora.dev"
train_corpus = "corpora.train"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.1
accumulate_gradient = 1
patience = 1600
max_epochs = 0
max_steps = 20000
eval_frequency = 200
frozen_components = []
annotating_components = []
before_to_disk = null
before_update = null

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

[training.batcher.size]
@schedules = "compounding.v1"
start = 100
stop = 1000
compound = 1.001
t = 0.0

[training.logger]
@Loggers = "spacy.ConsoleLogger.v1"
progress_bar = false

[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 = false
eps = 0.00000001
learn_rate = 0.001

[training.score_weights]
spans_sc_f = 1.0
spans_sc_p = 0.0
spans_sc_r = 0.0

[pretraining]

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

[initialize.components]

[initialize.tokenizer]

3. train_model.py

import spacy
import torch
from pathlib import Path
from spacy.cli.train import train

output_dir = Path("/Users/mgulenko/Dropbox/Research/KPIs/Python/spacy_spancat/training_data_span")
output_dir.mkdir(exist_ok=True)

train(
config_path="config.cfg",
output_path="training_data_span/output",
overrides={
"paths.train": str(output_dir / "train.spacy"),
"paths.dev": str(output_dir / "dev.spacy"),
"training.gpu_allocator": "pytorch",
}
)

I attached a 300 examples of our data, if that helps. Any help would be greatly appreciated!
sample_100.json