You can deploy this model and start earning money today!

qxswy/
modernbert-sports-ner-v2

Name: modernbert-sports-ner-v2
Availability: InStock

safetensors

modernbert

token-classification

named-entity-recognition

amazon-reviews

sports-outdoors

value-proposition

answerdotai/ModernBERT-large

finetune

apache-2.0

Views

No views yet

modernbert-sports-ner-v2

Domain-specific NER model for extracting value perception components from Sports & Outdoors product reviews. Fine-tuned as part of a methodology transfer experiment — applying a v1 pipeline originally developed for board game reviews to fitness equipment.

The label schema maps directly to a five-component consumer value model (adapted from Sheth-Newman-Gross, 1991):

Label	Value Component	Example
`PRODUCT`	Functional	"resistance bands", "handle", "latex"
`ATTRIBUTE`	Emotional	"durable", "snapped", "love it", "waste of money"
`PEOPLE`	Social	"seniors", "beginners", "physical therapist"
`EVENT`	Contextual	"home workout", "physical therapy", "travel"
`OTHER_PRODUCT`	Comparative	"TheraBand", "gym ones"

Performance

Metric	Value	Note
F1 macro (val)	0.75	entity-level, seqeval, n=100

Per-class breakdown was not computed for this checkpoint. The validation set (n=100) is small; true performance likely falls in the 0.65–0.80 range depending on class.

Based on the v1 model with the same architecture and training setup, expected relative performance: PRODUCT and EVENT are the strongest classes; ATTRIBUTE is the weakest due to its broad definition (spans both sentiments and physical qualities); OTHER_PRODUCT has very low support (20 entities in training).

Training

Base model: answerdotai/ModernBERT-large (395M params)
Dataset: 1,000 Amazon reviews (Sports & Outdoors), annotated synthetically via Gemini API
Annotation method: LLM generates span text + label; character offsets reconstructed locally via str.find() — 100% offset accuracy
Labels: BIO scheme, 5 entity types
Technique: R-Drop regularization (α=0.5) + Focal Loss (γ=2) + oversampling of rare classes (factor=3)
Hyperparameters: lr=3e-5, batch=4, max_length=256, weight_decay=0.01
Checkpoint: epoch 1 (best validation F1)

Limitations

Trained on synthetic annotations (Gemini API) — model learns LLM annotation patterns, not human patterns
Small validation set (n=100) means F1 estimate has high variance (±5–8%)
OTHER_PRODUCT has only 20 training entities — performance on this class is unreliable
ATTRIBUTE is a broad category covering both sentiments and physical qualities — this is a known bottleneck
Domain-specific: optimized for fitness/sports equipment reviews

Usage

python

from transformers import pipeline

pipe = pipeline(
    "token-classification",
    model="qxswy/modernbert-sports-ner-v2",
    aggregation_strategy="simple"
)

text = "Great bands for seniors doing home workouts, much better than TheraBand."
print(pipe(text))

Context

Part of a two-model NLP pipeline for value proposition analysis. Paired with qxswy/distilbert-sports-sentiment-v2 for sentiment classification.

See also: v1 models trained on board game reviews — qxswy/deberta-card-sentiment and qxswy/roberta-card-ner.