spacy-classes-finetune

Sentinel-D spaCy NER Model (Stage 1 — NVD Parsing)

Model Details

• Base Model: spaCy blank English (en_core_web_blank)
• Task: Named Entity Recognition (NER)
• Training Date: 2026-03-04T21:49:41.890810
• Framework: spaCy 3.x
• Training Data Size: 550 descriptions + 50-example test set
• Training Epochs: 20
• Dropout: 0.35

Custom NER Labels

1. VERSION_RANGE: Semantic version strings or version constraints (e.g., "1.2.3", "< 2.0.0")
2. API_SYMBOL: Method, class, or function names (e.g., "queryset.filter()", "X.509")
3. BREAKING_CHANGE: References to incompatible API changes or deprecations
4. FIX_ACTION: Specific remediation steps or upgrade instructions

Evaluation Metrics

Usage

import spacy

nlp = spacy.load("./spacy-nvd-ner-v1")

text = "OpenSSL versions before 1.1.1n contain a buffer overflow in the X.509 verifier."
doc = nlp(text)

for ent in doc.ents:
    print(f"{ent.text} -> {ent.label_}")
    # Output:
    # 1.1.1n -> VERSION_RANGE
    # X.509 -> API_SYMBOL

Installation

1. Extract the zip archive to your project directory
2. Load the model using spaCy:

   import spacy
   nlp = spacy.load("./spacy-nvd-ner-v1")
   

Architecture

The model consists of:

• Input Layer: Vectorized token representations
• Hidden Layer: Feed-forward network with 0.35 dropout
• Output Layer: 4-class NER tagger (softmax)

Training Configuration

• Optimizer: SGD
• Batch Size Range: 8-32 (compounding)
• Training Data: Real NVD descriptions auto-annotated with GLiNER teacher model
• Constraint: Exactly 50-example held-out test set (Master Document requirement)

Known Limitations

• Model trained on NVD descriptions only; may not generalize to other security domains
• Entity boundaries may not align perfectly with whitespace
• Requires English text input

License

MIT