durinn-data-filter 0.1.6

Risk-aware dataset decorator for Hugging Face datasets with automatic code detection and security risk labeling

Dataset Risk Decorator

A lightweight, plug-and-play risk annotation and filtering utility for Hugging Face datasets.

It annotates each sample with:

• risk_score
• is_problematic

while preserving full compatibility with the Hugging Face datasets API.

Designed for:

• Dataset auditing
• Safety filtering before fine-tuning
• Preference / DPO preprocessing
• Security research workflows

---

Maintainer

Durinn Research [e-mail] [e-mail] https://huggingface.co/durinn https://durinn.ai

---

Features

• Works with datasets.Dataset and datasets.DatasetDict
• Operates directly on loaded datasets
• Adds risk metadata without breaking downstream pipelines
• Optional row limiting for fast iteration
• Optional automatic filtering modes
• Trainer-compatible output

---

Installation

pip install durinn-data-filter

For local development:

python3.11 -m venv .venv
source .venv/bin/activate
pip install -e .

---

Quick Start

from datasets import load_dataset
from dataset_risk_decorator.core import risk_guard

ds = risk_guard(
    load_dataset("CyberNative/Code_Vulnerability_Security_DPO", split="train")
)

scores = ds["risk_score"]
print("rows:", len(ds))
print("min:", min(scores))
print("mean:", sum(scores) / len(scores))
print("max:", max(scores))

top_risky = ds.sort("risk_score", reverse=True).select(range(10))

for row in top_risky:
    print(row["risk_score"], row["chosen"][:200])

Each row now contains:

risk_score
is_problematic

---

Configuration

All configuration is optional and passed directly to risk_guard.

ds = risk_guard(
    load_dataset("CyberNative/Code_Vulnerability_Security_DPO", split="train"),
    threshold=0.5,
    filter_mode="none",
    max_rows=2000,
)

Parameters

Parameter	Description
threshold	Threshold used to set is_problematic
filter_mode	"none", "keep_safe", or "keep_problematic"
max_rows	Limit number of rows for faster iteration

---

Filtering Modes

filter_mode	Behavior
"none"	Annotate all rows
"keep_safe"	Keep only low-risk rows
"keep_problematic"	Keep only high-risk rows

Filtering happens after annotation.

---

Common Usage Patterns

Inspect Score Distribution

scores = ds["risk_score"]
print("rows:", len(ds))
print("min:", min(scores))
print("mean:", sum(scores) / len(scores))
print("max:", max(scores))

---

Train on Safe Samples Only

safe_ds = risk_guard(
    load_dataset("yahma/alpaca-cleaned", split="train"),
    filter_mode="keep_safe",
)

print(len(safe_ds))

---

Extract High-Risk Samples for Analysis

risky_ds = risk_guard(
    load_dataset("microsoft/Devign", split="train"),
    filter_mode="keep_problematic",
)

for row in risky_ds.select(range(10)):
    print(row["risk_score"], row.get("func", "")[:200])

---

DatasetDict Support

ds = risk_guard(
    load_dataset("yahma/alpaca-cleaned")
)

train = ds["train"]
test = ds["test"]

Each split is annotated independently.

---

Project Structure

dataset-risk-decorator/
├── dataset_risk_decorator/
│   ├── __init__.py
│   └── core.py
├── examples/
│   ├── default.py
│   └── with_config.py
├── pyproject.toml
└── README.md

---

Disclaimer

The current risk scorer is MVP-grade by design. It is intended to validate data pipelines and filtering mechanics prior to deploying a fully calibrated classifier.