EdyVision/pii-skills-ablation

PII Skills Ablation Benchmark

This repository provides the benchmark and experiment configuration for the ablation study described in:

"When Parametric Knowledge Wins: Evaluation Shapes Tool-Augmented PII Detection in Small Language Models"

The benchmark is a stratified sample of text with ground-truth PII spans aligned to PII-Codex canonical types. It is used to evaluate whether zero-shot prompting, documentation injection (+Docs), tool access (+Tool), or skills injection (+Skills) improve PII detection in small language models (7–9B parameters). The study also reports parametric baselines (few-shot and chain-of-thought prompting), a standalone PII-Codex detector run without a model, and two robustness controls that extend the analysis beyond the 7–9B class: a larger model (Qwen2.5-14B) and a full-precision (fp16/bf16) re-run. Predictions and per-sample scores for every condition, baseline, and control are released separately in the results dataset.

Dataset Summary

Overview

This dataset is an evaluation-only benchmark created to support controlled ablation experiments on agent context strategies for PII detection.

Rather than introducing new annotations, the dataset compiles and harmonizes multiple existing public PII datasets into a unified schema aligned with PII-Codex entity definitions. The resulting benchmark enables reproducible comparison across different prompting and agent configurations.

Key Characteristics

• Span-level PII annotations
• Canonical label alignment using PII-Codex types
• Stratified sampling to support controlled experimental comparison
• Designed for evaluation (not training)

Language

English

Dataset Structure

• test: Main benchmark split. The dataset is dehydrated: it stores sample identifiers and ground-truth labels only; raw text is not included and must be resolved from the original source datasets using source and original_index (see code repo).

Additional assets in this repo (under config/):

• config/config.yaml: Experiment and sampling configuration.
• config/pii_label_to_piicodex.json: Mapping from source PII labels to PII-Codex canonical types (applied symmetrically to predictions and ground truth).
• config/prompts/v1/: Prompt templates for every condition and baseline used in the study: zero_shot.txt, with_docs.txt, with_tools.txt, with_skills.txt, few_shot.txt, and cot.txt. The directory also contains skill_workflow.txt, the PII-detection Skill document (SKILL.md payload) returned to the model by view_skill("pii-detection") in the +Skills condition.

The run configurations for the few-shot/chain-of-thought baselines and the scale/precision controls (rr_baselines.yaml, rr_scaling.yaml, rr_fp16.yaml) are in the code repository under config/.

Data Fields

Field	Type	Description
id	string	Unique sample identifier (e.g. ai4_123, nvidia_456, gretel_789).
source	string	Origin dataset: ai4privacy, nvidia_nemotron_pii, or gretel_pii_masking.
original_index	int	Index of the sample in the source dataset; used to load text from the source at run time.
ground_truth	list	Original source PII type labels per entity: list of {"type": "<label>"}.
pii_codex_ground_truth	list	Same entities mapped to PII-Codex canonical type names (e.g. PERSON, EMAIL_ADDRESS) for evaluation.

Text is not stored in this dataset; the evaluation pipeline loads it from the corresponding HuggingFace source dataset using source and original_index.

Span structure (source data and model predictions): Where span-level offsets are used (e.g. in source datasets or in the results repo), each predicted span has:

{
  "type": "EMAIL_ADDRESS",
  "text": "john.smith@acme.com",
  "start": 22,
  "end": 41
}

type is a PII-Codex type name, text is the exact span, and start/end are character offsets into the text. The benchmark itself stores type-only ground truth (ground_truth / pii_codex_ground_truth), not character spans.

Intended Use

• Primary: Research and reproducibility for the ablation study (evaluating agent configurations on PII detection).
• Secondary: Benchmarking PII detection or NER models on a consistent, aligned schema. Not intended for production deployment without further validation.
• Out-of-Scope Uses: Not intended for production deployment without further validation. Not designed as a comprehensive general-purpose PII dataset.

Ethical Considerations

• Source datasets may include synthetic or publicly released PII examples.
• Alignment transformations may affect semantic boundaries.
• Users should not assume that all examples are free of sensitive content risks.
• Responsible handling of PII-related data is required.

Citations

If you use this benchmark or config, please cite the paper and, where applicable, the dataset:

@article{rosado2026context,
  title   = {When Parametric Knowledge Wins: Evaluation Shapes Tool-Augmented PII Detection in Small Language Models},
  author  = {Rosado, Eidan J.},
  year    = {2026},
  note    = {Benchmark: \url{https://huggingface.co/datasets/EdyVision/pii-skills-ablation}}
}

Source Dataset Citations

This benchmark compiles and aligns samples from the following datasets. If you use this benchmark, please also cite the original datasets:

AI4Privacy PII-Masking-300K

@misc{ai4privacy_2024,
    author       = { Ai4Privacy },
    title        = { pii-masking-300k (Revision 86db63b) },
    year         = 2024,
    url          = { https://huggingface.co/datasets/ai4privacy/pii-masking-300k },
    doi          = { 10.57967/hf/1995 },
    publisher    = { Hugging Face }
}

NVIDIA Nemotron-PII

@dataset{nemotron-pii,
  author = {Amy Steier and Andre Manoel and Alexa Haushalter and Maarten Van Segbroeck},
  title = {Nemotron-PII: Synthesized Data for Privacy-Preserving AI},
  year = {2025},
  publisher = {NVIDIA},
  url = {https://huggingface.co/datasets/nvidia/Nemotron-PII}
}

Gretel PII Masking (English)

@dataset{gretel-pii-docs-en-v1,
  author       = {Gretel AI},
  title        = {GLiNER Models for PII Detection through Fine-Tuning on Gretel-Generated Synthetic Documents},
  year         = {2024},
  month        = {10},
  publisher    = {Gretel},
}

Limitations

• Ground truth is derived from source datasets and aligned to PII-Codex; alignment choices can affect edge cases and rare types.
• The benchmark is sampled for experimental control (stratification by PII types and text length); it is not a general-purpose PII benchmark for all domains.
• Prompt and config versions (e.g. v1) are fixed for the reported study; other versions may be added later.

License

See the dataset repository and the linked paper/code repository for license terms. Source datasets (AI4Privacy, NVIDIA, Gretel) have their own licenses; compilers must comply with those when redistributing or deriving data.

Related Resources

• Scored results (predictions and metrics): EdyVision/pii-skills-ablation-results — splits for the main and pilot ablations plus the few-shot/chain-of-thought baselines, standalone detector, 14B scale control, and full-precision control.
• Code and notebooks: EdyVision/pii-skills-ablation-study