CoolFace
Datasetpublic

tunedtensor/email-triage-v1

Email Triage v1 This dataset contains 1,740 unique email-triage examples produced through Tuned Tensor labeling and hardening workflows. It is the public v1 dataset, de-duplicated from a 2,038-row weighted fine-tuning dataset; 298 intentional weighting duplicates were removed for easier reuse. The task is operational inbox triage, not security-risk classification. Each row asks a model to label one email-like message and return strict JSON with triage, priority, should_process… See the full description on the dataset page: https://huggingface.co/datasets/tunedtensor/email-triage-v1.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes64downloads
Dataset Card

Email Triage v1

This dataset contains 1,740 unique email-triage examples produced through Tuned Tensor labeling and hardening workflows. It is the public v1 dataset, de-duplicated from a 2,038-row weighted fine-tuning dataset; 298 intentional weighting duplicates were removed for easier reuse.

The task is operational inbox triage, not security-risk classification. Each row asks a model to label one email-like message and return strict JSON with triage, priority, should_process, confidence, summary, and reason.

Files

  • —data/train.jsonl: 1,740 unique {input, output} training examples.
  • —data/targeted_boundary_645_with_source.jsonl: latest targeted boundary-label slice with parsed labels and source provenance.
  • —metadata/build-summary.json: counts and Tuned Tensor IDs for this public package.
  • —metadata/source-2038-weighted-build-summary.json: build summary for the source weighted training dataset.
  • —metadata/targeted-boundary-645-summary.json: build summary for the targeted 645-label slice.
  • —metadata/tunedtensor-spec-snapshot.json: behavior spec snapshot used for this labeling/fine-tuning workflow.

Label Schema

The output field is a compact JSON string with exactly:

  • —triage: one of reply, archive, escalate, ignore, or review.
  • —priority: one of low, normal, high, or critical.
  • —should_process: boolean. true for reply, escalate, or review; false for archive or ignore.
  • —confidence: numeric confidence from 0 to 1.
  • —summary: concise neutral content summary.
  • —reason: short explanation for the triage decision.

Counts

Rows: 1,740

Source weighted training rows: 2,038

Removed weighting duplicates: 298

Triage distribution:

  • —escalate: 554
  • —reply: 471
  • —ignore: 300
  • —archive: 268
  • —review: 147

Priority distribution:

  • —low: 576
  • —critical: 444
  • —high: 389
  • —normal: 331

Usage

python
from datasets import load_dataset

ds = load_dataset("tunedtensor/email-triage-v1")
print(ds["train"][0]["input"])
print(ds["train"][0]["output"])

targeted = load_dataset(
    "tunedtensor/email-triage-v1",
    "targeted_boundary_enriched",
)
print(targeted["train"][0]["label"])

Provenance

  • —Labeling/fine-tuning system: Tuned Tensor
  • —Behavior spec ID: 27cfdfb2-28af-4c77-ac32-052c3ab6194e
  • —Source weighted Tuned Tensor dataset ID: 98c0ddc5-0a1e-4700-b00c-e4f8432f2650
  • —Targeted labeling job ID: 3b52e908-8b82-4d64-8704-01c8e3171f24
  • —Targeted promoted dataset ID: fa4435b6

The dataset combines unique examples from a previous hardening set plus a new 645-row targeted boundary-label slice.

The targeted slice was sampled from:

  • —weijianzhg/email-triage-action-seed synthetic email-triage seed data. The source dataset declares Apache-2.0 license metadata.
  • —SetFit/enron_spam Enron spam/ham data. The Hugging Face dataset card did not declare license metadata at preparation time.

Limitations

Labels were generated automatically and should be reviewed before production use. The dataset is intentionally focused on inbox-triage behavior and should not be treated as a complete email-security or phishing-risk classifier. Some source rows are public Enron-derived emails and may contain dated business context or noisy text.