README.md

---
license: other
task_categories:
- text-classification
language:
- en
tags:
- legal
- enron
- email
- correspondence
- document-classification
- evaluation
- deduplicated
- ground-truth
- sentiment-analysis
- topic-classification
- llm-mailroom
pretty_name: "Enron Correspondence Deduplicated (Enriched GT, Agent-Blind Default)"
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: parquet/blind/train/train-*.parquet
  - split: test
    path: parquet/blind/test/test-*.parquet
- config_name: ground_truth
  data_files:
  - split: train
    path: parquet/ground_truth/train/train-*.parquet
  - split: test
    path: parquet/ground_truth/test/test-*.parquet
---

# Enron Correspondence Deduplicated (Enriched GT, Agent-Blind Default)

The **deduplicated, ground-truth-enriched** companion to
[`Lucius-Morningstar/enron-correspondence`](https://huggingface.co/datasets/Lucius-Morningstar/enron-correspondence):
exact-duplicate bodies removed from the cleaned CMU Enron corpus (517,390
rows in -> **247,523 unique-text rows out**, 269,867 duplicates dropped;
first occurrence wins on maildir-path order; empty bodies never deduped against
each other).

## ⚠️ Two-config layout: agents get NO answers by default

This dataset ships TWO configs. **`default` is agent-blind** — it carries the
email content plus routing metadata and ZERO ground-truth columns. The answer
keys live in the separate `ground_truth` config, keyed 1:1 on `filename`.

```python
from datasets import load_dataset

# what a sorting/extraction agent may see:
blind = load_dataset("Lucius-Morningstar/enron-correspondence-dedup")

# what the scorer joins against (explicit opt-in):
gt = load_dataset("Lucius-Morningstar/enron-correspondence-dedup",
                  "ground_truth", split="test")
```

| config | splits | columns |
|---|---|---|
| `default` | train 222,572 / test 24,951 | filename, subject, text, split, metadata |
| `ground_truth` | train 222,572 / test 24,951 | filename, expected, expected_subclass, label_evidence, content_topic, topic_evidence, sentiment_score, sentiment_label, sentiment_evidence, split |

Ground truth is hidden from the default config so automated agents cannot be
tipped off; humans can still audit every label in the viewer by switching to
the `ground_truth` config. This is separation of concerns, NOT encryption —
the Hub is public and any deliberate download can fetch both configs.

## File layout

Both configs point at pre-sharded **parquet** (`parquet/<config>/<split>/*.parquet`,
zstd). The original `blind/*.jsonl` and `ground_truth/*.jsonl` files remain in
the repo unchanged — row-for-row identical to the parquet shards (same order,
same schema; verified 2026-08-23) — for pipelines that prefer line-delimited
JSON. The single-file 548 MB `blind/train.jsonl` crashed the Dataset Viewer's
conversion worker (`JobManagerCrashedError`); serving parquet directly removes
that conversion step entirely.

## Ground-truth dimensions

1. **doc_type / subclass** (`expected`, `expected_subclass`,
   `label_evidence`) — heuristic form taxonomy from the shared
   [`correspondence_subclasses`](https://github.com/Exios66/Enron-Evaluation-Environment)
   labeler: attorney_demand, demand, email, letter, meeting_request, memo, notice, press_release.
2. **content_topic** (`content_topic`, `topic_evidence`) — WHAT the message
   body is about: an 11-key priority-scored marker taxonomy
   (`content_topics.py`): legal_contracts, regulatory, finance_earnings, energy_market, hr_personnel, it_systems, travel_logistics, marketing_clients, announcements, scheduling.
3. **sentiment** (`sentiment_score` ∈ [-1, 1], `sentiment_label` ∈
   negative/neutral/positive, `sentiment_evidence`) — deterministic lexicon
   polarity over the subject + forwarded-tail-stripped body
   (`sentiment_scorer.py`), negation/intensifier-aware, politeness-formula
   controlled.

All three dimensions are HEURISTIC ground truth (deterministic pure functions,
human-reviewed via spot checks where noted) — not hand annotations. Honest
gaps: single-topic assignment for multi-topic emails; head-window scanning
(~2000 chars); lexicon sentiment cannot read sarcasm or long-range context —
treat scores as weak labels/routing priors. Attorney detection relies on
domain/name lists; `voicemail` cannot occur in this text-only corpus.

## Splits

Per-row `split` follows the family rule `md5(filename) % 10 == 0 -> test`
(~10%), recomputed and asserted row-by-row at build time. Filename-keyed, so
dedup/enrichment cannot change any surviving row's split. Coverage: train
222,572 / test 24,951.

## Provenance

Built by [`llm-entity-extraction`](https://github.com/Exios66/llm-entity-extraction)
`scripts/datasets/publish_enron_correspondence_dedup.py` (KANBAN-079,
2026-08-23T18:52:37+00:00) from the sha256-verified full-corpus export (LFS
`0554a5973935…`). Labelers: Enron-Evaluation-Environment `scripts/`
(`correspondence_subclasses.py`, `content_topics.py`,
`sentiment_scorer.py`). Source: CMU Enron Email Dataset (cleaned maildir);
dedup rule `scripts/dedupe.py::body_hash`. Research-use license — treat
personally identifying content accordingly.