danielrosehill/pseudo-personalised-outreach

Pseudo-personalised outreach

Real unsolicited emails sent to one person, labelled genuine or pseudo_personalised, with every sender identity replaced by a plausible fake.

• Code, spec and labelling rules: https://github.com/danielrosehill/Pseudo-Personalised-Spam-Classifier
• Labels and subtypes: docs/spec.md#labels
• How identities were replaced: docs/pseudonymisation.md

Intended use

Training and evaluating a small, cheap classifier (SetFit / ModernBERT class, well under 1B parameters) that labels inbound cold outreach in a personal inbox:

• genuine: a real person put thought into contacting the recipient about something of theirs, and the ask follows from it. This includes sales pitches; the question is genuineness, not whether the mail is wanted.
• pseudo_personalised: mass or automated outreach dressed up as a personal note. This includes LLM-researched messages whose "personal" detail is accurate but decorative, where the ask would be the same for anyone.

The intended deployment only adds a label to mail. It is not meant to delete, block or auto-reply. Classic spam, phishing and newsletters are out of scope; a rule-based pre-filter handles them before the model runs.

Not for: identifying or contacting senders, judging people or companies, or general spam filtering.

Data

data/train.jsonl, one JSON object per line. The schema is schema/email-record.schema.json.

Field	Notes
subject, body_text	Pseudonymised; style, typos and formatting preserved
label	genuine / pseudo_personalised
subtype	Annotation-only sub-category (provisional in this release)
campaign_group	Rows from the same sender share it; use for grouped CV splits
channel, received_month	Arrival route and YYYY-MM
recipient	Which of the recipient's addresses got it, as {kind, alias} (public contact, GitHub, a platform-only alias…). A strong signal: harvested platform aliases only ever received pseudo-personalised mail in the seed set
header_features	Booleans and counts derived from the raw headers (unsubscribe header, tracking links, link count, reply-to mismatch, sender provider, HTML-only)
judge	A small LLM's blind label, confidence, signal tags and short rationale: why the message reads as it does. Metadata for auditing and a zero-shot baseline; not the training target

Counts in this release: {"genuine": 4, "pseudo_personalised": 14}. The recipient ("Daniel") and his own public projects are kept, because the personalisation hook is the signal.

Optional views (plain/HTML/headers) will appear as raw MIME becomes available; see the spec's Dataset views.

How senders were protected (methodology)

This dataset is built from real mail, so the senders were protected before publication. The goal is to avoid naming and shaming anyone who sent unsolicited mail. The recipient side is deliberately kept: Daniel Rosehill, his companies and his public projects appear as written, because the "personal" hook pointing at his work is what the classifier must learn to judge.

It is a custom, LLM-assisted transformation, not an anonymisation model:

1. A small LLM (Claude Haiku) rewrote each message by fixed written rules. Every sender name, company, address, domain, URL, phone number, handle and ID became a plausible fake of the same shape (jason.m@acme.ai → owen.h@brindlepath.ai). Nothing became a [PLACEHOLDER], because placeholders would teach the model a token that real mail never contains. Wording, typos, layout, prices and dates are unchanged. Fakes for both classes come from the same pool.
2. A deterministic script checked every rewrite against its original for leftover names, addresses, domains, URL hosts and digit runs.
3. A second reviewer read every row for prose-only names and for unique combinations of facts that could identify a sender even without names, and rewrote those.
4. Anything sent in confidence was withheld. The real→fake mapping was discarded.
5. Raw headers were reduced to derived features (header_features). Which of the recipient's addresses got the mail is kept as a category (recipient).

Full rules and the procedure as run: pseudonymisation.md (copied from the repo). If you recognise your own message, open an issue on the GitHub repo and the row will be removed.

Status

Seed release, 2026-10-20: 18 rows labelled by the recipient. Too small to train on alone; it grows as annotation continues (target ~100 for the proof of concept).