Releases: Erox-02/humming-bird-v3
Release list
Hbp100 v3.0.0
HBP100 v3.0.0
Hbp100 detects sensitive information locally, uses a contextual machine learning policy engine to decide what should be masked, replaces sensitive values with placeholders, and restores them after external processing.
What's new in v3.0.0
Full Rust runtime
Hbp100 v3 removes the Python runtime from the production stack.
The entire production pipeline now runs natively in Rust:
- Entity extraction
- Contextual feature extraction
- Machine learning inference
- Privacy decisions
- Placeholder generation
- Metadata handling
- Restoration
No Python runtime is required.
Native LightGBM inference
The privacy policy engine now uses LightGBM directly from Rust through lightgbm3.
The trained model is distributed as:
assets/hbp100-v3.lgb
The model can be loaded and executed directly by the Rust runtime without Python, pickle files, scikit-learn, or a Python ML bridge.
Context-aware privacy decisions
Hbp100 keeps the surrounding context when making privacy decisions.
The model does not simply ask:
"Does this look like a name?"
It evaluates the entity together with its surrounding text, entity type, and optional intent.
This allows the policy engine to decide whether masking an entity would unnecessarily destroy useful context.
New v3 ML pipeline
The v3 model was trained entirely from the Rust training pipeline.
Current dataset:
- 4,612 entity samples
- 1,239 features
Current held-out evaluation:
| Metric | Score |
|---|---|
| Accuracy | 91.34% |
| Precision | 95.57% |
| Recall | 84.59% |
| F1 | 89.75% |
The model is designed to balance privacy with contextual usefulness rather than blindly masking every detected entity.
Release
Current github release contains /dev/test.rs , the crates release doesnt contain that .
The crates release can be viewed on :
https://crates.io/crates/hbp100/3.0.0
also can be installed by
cargo install hbp100
to use the github release download the source code of this release and run
cargo --build
Hbp100 v2.2.0
HBP100 v2.2.0
Contextual privacy firewall using machine learning for intelligent PII masking.
Features
~ LightGBM + TF-IDF privacy engine
~ Context-aware masking decisions
~ Placeholder generation and validation
~ Metadata vault for reversible masking
~ Restore API
~ Lightweight package (~590 KB wheel)
~ Sub-millisecond average latency
~ MIT licensed and open source
Improvements in v2.2
~ Added multi prompt support
~ Optimized for servers
~ Fixed Bug on long paragraph extractions
~ Better address extraction
~ Better phone number normalization
~ Reduced duplicate entity detection
~ More robust masking pipeline
Example
from hbp100 import mask, restore
text = "My email is john@example.com"
masked = mask(text)
# My email is [EMAIL_1]
response = llm(masked)
final = restore(response)
# My email is john@example.comApplications
~ Healthcare
~ Insurance documents
~ OCR pipelines
~ AI assistants
~ Customer support systems
~ Privacy-preserving LLM workflows
HBP100 v2.0.0
HBP100 v2.0.0
Contextual privacy firewall using machine learning for intelligent PII masking.
Features
~ LightGBM + TF-IDF privacy engine
~ Context-aware masking decisions
~ Placeholder generation and validation
~ Metadata vault for reversible masking
~ Restore API
~ Lightweight package (~590 KB wheel)
~ Sub-millisecond average latency
~ MIT licensed and open source
Improvements in v2
~ Improved entity overlap handling
~ Fixed nested replacement issues
~ Better address extraction
~ Better phone number normalization
~ Reduced duplicate entity detection
~ More robust masking pipeline
Example
from hbp100 import mask, restore
text = "My email is john@example.com"
masked = mask(text)
# My email is [EMAIL_1]
response = llm(masked)
final = restore(response)
# My email is john@example.comApplications
~ Healthcare
~ Insurance documents
~ OCR pipelines
~ AI assistants
~ Customer support systems
~ Privacy-preserving LLM workflows