build-small-hackathon/pakistan-notice-helper-traces

Pakistan Notice Helper Privacy-Safe Traces

Purpose

This dataset contains compact, deterministic metadata about Pakistan Notice Helper scam-check requests. It does not contain hidden model reasoning or autonomous-agent trajectories.

The application uses a Modal-hosted Qwen model for normal assessments. Creating a trace never makes an additional AI model call. Traces only observe the existing request path and convert it into allow-listed categories, booleans, and fixed descriptions. For image submissions, the existing assessment's explanation and red flags may be inspected transiently for this mapping, but their text is not stored. New model responses also include enum-only category and tactic hints. These hints are treated as untrusted: the trace mapper keeps only values supported by the explanation or red flags and falls back to its deterministic English/Urdu evidence rules when they disagree.

Fields

• Trace identity: random trace ID and UTC timestamp
• input: aggressively redacted, length-limited text for text submissions, or a fixed-template image: ... description
• input_category: deterministic category such as courier, bank, or FBR
• urgency: deterministic boolean urgency signal
• result_summary: deterministic summary of the mapped pattern, whether it is known or unclassified, and why the risk label matters
• scam_tactics: readable comma-separated tactics
• Flat result columns such as risk_label and reply_draft_policy

Every dataset cell is a scalar string, number, or boolean. No column contains a dictionary or nested object, which keeps the Hugging Face table easy to read. trace_id is the first serialized column, and all detected boolean signals are combined into the single scam_tactics category column.

Records do not contain pipeline steps, cache fields, app commits, failure details, request source, schema versions, size buckets, language hints, or Modal metadata.

Privacy

The dataset never stores:

• Raw message text
• Screenshots, image bytes, or base64
• URLs, phone numbers, CNICs, names, addresses, account/card numbers, or tracking numbers
• Model explanations, red flags, safe-step text, reply drafts, or raw model output
• Exceptions, credentials, tokens, or endpoint headers

Text traces store an aggressively redacted form capped at 500 characters. Regexes remove common URLs, emails, phones, CNICs, card/account numbers, credentials, addresses, tracking IDs, long numbers, and title-case names or entities. Regex redaction cannot guarantee removal of every possible identifier, so users should opt out when submitting sensitive content.

Images store only fixed descriptions. Screenshots, OCR text, model explanations, and model red flags are not stored. Users see a checked trace disclosure in the app and may opt out before each request.

When image analysis fails before an assessment is available, new records use image: Assessment unavailable. This keeps service failures separate from successful assessments whose content is genuinely unclassified.

Provenance

Seed traces represent the six public examples bundled with Pakistan Notice Helper. Runtime traces may represent successful, rejected, or failed requests. Trace generation itself does not invoke the model.

The seed rows are illustrative examples, not an evaluation split. All six currently have the Likely scam label, so they must not be used to estimate class balance, accuracy, recall, or real-world scam prevalence.

Runtime rows are an operational log and intentionally preserve repeated requests. Consequently, repeated examples, unclassified image descriptions, and incomplete none assessments may be common. For training or evaluation, create a separate curated split that:

• excludes risk_label: none
• reviews or excludes unclassified image rows
• deduplicates on the privacy-safe input and result columns
• uses a task-appropriate class-balancing strategy

The source repository includes traces/scripts/analyze_trace_dataset.py for schema validation and a reproducible summary of these quality indicators.

Limitations

• Regex signals and category detection are approximate.
• Runtime frequencies may reflect testing or repeated usage, not population prevalence.
• Regex redaction may miss unusual personal or confidential information.
• Novelty is not researched against external threat-intelligence sources.
• The dataset cannot reproduce original messages or screenshots.
• A risk label is safety guidance, not official verification.

Links

• App: https://huggingface.co/spaces/build-small-hackathon/pakistan-notice-helper
• Source: https://github.com/kingabzpro/pakistan-notice-helper

License

CC BY 4.0.