familyos-ultrabert 4.0.3

FamilyOS UltraBERT v4 - Multi-task NLP with GlobalPointer NER

FamilyOS UltraBERT v2

High-performance multi-task NLP model for family communication analysis.

Built on ModernBERT architecture, UltraBERT delivers 12 NLP capabilities in a single unified model - extract sentiment, emotions, safety signals, entities, and more from text with state-of-the-art accuracy.

Features

• 12 NLP Capabilities in one unified model
• PyTorch & ONNX backends for flexible deployment
• < 15ms latency for 6 capabilities on GPU
• Single encoder pass for multi-capability inference
• 768-dim sentence embeddings for semantic search
• 155M parameters - Optimized ModernBERT architecture

Installation

From GitHub (recommended)

# Clone the full repository
git clone https://github.com/Pkansagra-hub/Family_osModernBERT.git
cd Family_osModernBERT

# Install the package with PyTorch backend
pip install ./familyos_ultrabert[pytorch]

# Or with ONNX backend
pip install ./familyos_ultrabert[onnx]

# Or both
pip install ./familyos_ultrabert[all]

From PyPI (coming soon)

# Basic installation
pip install familyos-ultrabert

# With PyTorch backend (recommended for GPU)
pip install familyos-ultrabert[pytorch]

# With ONNX backend (recommended for CPU)
pip install familyos-ultrabert[onnx]

# Full installation (both backends)
pip install familyos-ultrabert[all]

Note: The PyTorch backend requires the full repository for model architecture code. The ONNX backend is fully standalone.

Quick Start

from familyos_ultrabert import UltraBERT

# Load model (auto-selects best backend)
model = UltraBERT.load()

# Analyze text with multiple capabilities
result = model.analyze(
    "Mom picked up Panda from school today!",
    capabilities=["sentiment", "ner_family", "safety_familyos", "emotions"]
)

# Access results
print(result["sentiment"])
# {'prediction': 'positive', 'confidence': 0.89, 'scores': {...}}

print(result["ner_family"]["entities"])
# [{'text': 'Mom', 'label': 'KINSHIP'}, {'text': 'Panda', 'label': 'NICKNAME'}]

print(result["safety_familyos"])
# {'band': 'GREEN', 'confidence': 0.98, 'probabilities': {...}}

print(result["emotions"]["predictions"])
# ['joy', 'caring', 'togetherness']

Capabilities

Capability	Type	Description
sentiment	Classification	5-class sentiment (very_negative to very_positive)
emotions	Multi-label	44 emotions including family-specific feelings
safety_familyos	Classification	Safety bands: GREEN, AMBER, RED, CRISIS
safety_generic	Multi-label	8 toxicity types (Jigsaw-style)
intent	Classification	8 user intents (log_memory, query_memory, etc.)
ingress	Classification	12 domain categories (DIARY, TASK, HEALTH, etc.)
ner_family	Token	Family entities (KINSHIP, NICKNAME, PET, etc.)
ner_general	Token	General NER (PER, ORG, LOC, DATE, etc.)
temporal	Token	Temporal expressions (DATE_ABS, DATE_REL, DURATION)
relation	Multi-label	15 relationship types (parent_of, spouse_of, etc.)
nli	Classification	Natural language inference
embedding	Vector	768-dim sentence embeddings

Convenience Methods

# Sentiment
sentiment = model.get_sentiment("I love this!")
print(sentiment["prediction"])  # "very_positive"

# Emotions
emotions = model.get_emotions("So excited for the trip!")
print(emotions)  # ["excitement", "joy", "anticipation"]

# Safety check
band = model.get_safety_band("Having a great day!")
print(band)  # "GREEN"

# Entity extraction
entities = model.get_entities("Mom and Dad took the kids to grandma's house")
print(entities)
# [{'text': 'Mom', 'label': 'KINSHIP'}, {'text': 'Dad', 'label': 'KINSHIP'}, ...]

# Embeddings
embedding = model.get_embedding("Sample text for embedding")
print(len(embedding))  # 768

Backend Selection

# Auto-detect (uses GPU if available)
model = UltraBERT.load()

# Force PyTorch on GPU (best for multi-capability)
model = UltraBERT.load(backend="pytorch", device="cuda")

# Force ONNX on CPU (best for single-capability, deployment)
model = UltraBERT.load(backend="onnx", device="cpu")

# Custom model path
model = UltraBERT.load(model_path="/path/to/weights")

Model Architecture

Component	Details
Base Model	ModernBERT-base (22 layers, 768 hidden)
Parameters	155M total
Encoder	Shared transformer backbone
Heads	12 specialized task heads
Optimization	15% magnitude pruning
Quantization	Dynamic INT8 (ONNX)

Benchmarks (RTX 4090)

Running the built-in benchmark suites

Benchmarks ship with the package and can be run either as a module:

python -m familyos_ultrabert.benchmarks --suite api,regression

or via the console script (when installed):

ultrabert-benchmark --suite api,regression

Reporting formats:

python -m familyos_ultrabert.benchmarks --suite api,regression --format text
python -m familyos_ultrabert.benchmarks --suite api,regression --format json --output benchmark_report.json
python -m familyos_ultrabert.benchmarks --suite api,regression --format markdown --output benchmark_report.md

For a faster smoke run:

python -m familyos_ultrabert.benchmarks --quick

Standard profiles (recommended for CI):

python -m familyos_ultrabert.benchmarks --profile smoke
python -m familyos_ultrabert.benchmarks --profile full

Baseline drift tracking (compare against last-known-good per environment key and update baseline):

python -m familyos_ultrabert.benchmarks --profile smoke --baseline --format json --output benchmark_report.json

Task Performance

Task	Metric	Score
safety_familyos	Accuracy	96.20%
intent	Actionable Rate	96.58%
emotions	Hit Rate	88.30%
sentiment	Direction Accuracy	88.10%
ner_family	F1	87.71%
temporal	F1	87.17%
Weighted Average		89.60%

Latency

Scenario	Latency	Throughput
1 capability	8 ms	-
6 capabilities	14 ms	-
12 capabilities	14 ms	71 samples/sec
Embedding query	13 ms	1,921 emb/sec

License

Proprietary - All Rights Reserved. See LICENSE for details.