Datasets:

AriaAICompany
/

phish-messages

n_test int64	split string	never_outputs_safe bool	evidence_link_rate float64	hard_negative_suspicious_rate float64	decision_counts dict	methods dict	by_language dict	limitations list
96	campaign-held-out	true	1	0	{ "suspicious": 48, "needs_review": 16, "insufficient_evidence": 32 }	{ "rule-structural": { "pr_auc": 1, "brier": 0.10375000000000001, "recall_at_fpr_0_05": 1, "recall_at_fpr_0_10": 1, "mean_score": 0.34166666666666673, "pr_curve_points": 6 }, "tfidf-lr": { "pr_auc": 1, "brier": 0.14890835989583331, "recall_at_fpr_0_05": 1, "recall_at_fpr_0_...	{ "en": { "n": 64, "pr_auc_lightgbm": 1, "recall_at_fpr_0_10": 1 }, "fa": { "n": 32, "pr_auc_lightgbm": 1, "recall_at_fpr_0_10": 1 } }	[ "Synthetic fixtures only. Lab PR-AUC is not operational phishing accuracy.", "Copied Authentication-Results headers are not treated as proof.", "URLs are parsed and defanged; they are never fetched.", "A similar logo or a new domain is not enough for a malice verdict." ]

Phish synthetic messages

256 synthetic Persian and English messages for the Phish review demo. Seed 3.

Organization dataset, model, collection, and static card are public. Live Gradio is created by scripts/publish.py. This is fixture data (level 1). It does not prove operational phishing accuracy.

Files

data/messages.jsonl
data/splits.json
data/evaluation.json
data/sample_preview.json
data/eml/*.eml
data/protocol.md

Splits

Split is by campaign, sender group, and template, not random rows of the same lure.

development: 10 campaigns × 16 variants = 160
test: 6 campaigns × 16 variants = 96
interactive: six *-00 fixtures used in the Space

Phishing, hard-negative invoices/login notices, and insufficient-evidence stubs are all present. Hard negatives keep aligned From/Reply-To/href domains.