DeL-TaiseiOzaki/enronhop-timesplit_2001_0418

EnronHop — Time Split (cutoff: 2001-04-18)

Time-based re-split of the miluki/enronhop multi-hop Q&A benchmark, enabling temporal generalization evaluation.

Split policy

Items from the combined miluki/enronhop train+test pool (1,947 items total) are re-partitioned at cutoff 2001-04-18 00:00:00:

• train (1,360 items, 69.85%): max(evidence_dates) < cutoff
• test (587 items, 30.15%): max(evidence_dates) >= cutoff

evidence_dates are looked up per item by joining evidence_mails (mail_ids) with the date field in miluki/enronhop-corpus/emails.jsonl (the kaminski-v custodian subset, 11,062 emails).

Files

• train.jsonl — 1,360 items
• test.jsonl — 587 items
• manifest.json — detailed provenance

Item schema matches miluki/enronhop exactly (same columns, just re-partitioned).

Companion corpus (for leak-free training)

A date-filtered corpus subset is published alongside for proper temporal evaluation:

File	Emails	Use
DeL-TaiseiOzaki/enronhop-corpus/emails.kaminski_v.jsonl	11,062	Full search index (for test-time eval)
DeL-TaiseiOzaki/enronhop-corpus/emails.kaminski_v.pre_cutoff.jsonl	6,784	Pre-cutoff search index (for train-time)

During training, point the agent's search index to the pre_cutoff file to prevent future-data leakage. For test evaluation, use the full kaminski_v corpus.

Usage with enronhop_env

No environment code changes are needed — enronhop_env.load_environment() already accepts train_data_file, test_data_file, corpus_data_file, and db_path as arguments. Override them in your Prime RL config or direct Python call:

Prime RL config (TOML)

# Training env — leak-free pre-cutoff corpus
[[env]]
id = "miluki/enronhop_env"
args = { max_turns = 15, judge_model = "gpt-5-nano",
  train_data_file  = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/train.jsonl",
  test_data_file   = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/test.jsonl",
  corpus_data_file = "hf://datasets/DeL-TaiseiOzaki/enronhop-corpus/emails.kaminski_v.pre_cutoff.jsonl",
  db_path          = "~/.cache/enronhop_env/enronhop_corpus.pre_cutoff.db" }

# Eval env — full corpus so test items' future evidence is reachable
[[eval.env]]
id = "miluki/enronhop_env"
args = { max_turns = 10, judge_model = "gpt-5-nano",
  train_data_file  = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/train.jsonl",
  test_data_file   = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/test.jsonl",
  corpus_data_file = "hf://datasets/DeL-TaiseiOzaki/enronhop-corpus/emails.kaminski_v.jsonl",
  db_path          = "~/.cache/enronhop_env/enronhop_corpus.full.db" }

Important: always use distinct db_path values for the train and eval envs, or the SQLite cache built from the first corpus will be reused by the second one (silent corpus mismatch).

Direct Python

from enronhop_env import load_environment

env = load_environment(
    train_data_file  = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/train.jsonl",
    test_data_file   = "hf://datasets/DeL-TaiseiOzaki/enronhop-timesplit_2001_0418/test.jsonl",
    corpus_data_file = "hf://datasets/DeL-TaiseiOzaki/enronhop-corpus/emails.kaminski_v.pre_cutoff.jsonl",
    db_path          = "~/.cache/enronhop_env/enronhop_corpus.pre_cutoff.db",
)

Hop distribution

hop	train	test	test % of hop
1	465	124	21.1%
2	531	235	30.7%
3	244	126	34.1%
4	120	102	45.9%
total	1,360	587	30.15%

Multi-hop items skew toward test because their evidence can span larger date ranges, so max(evidence_dates) lands post-cutoff more often.

Integrity

• All 1,947 items annotated successfully (100% evidence coverage in corpus).
• Train items: 0 evidence mails missing from pre_cutoff corpus (full access at train time).
• Test items: all 587 have at least one evidence mail removed by the pre_cutoff filter (319 with ALL evidence post-cutoff, 268 partial).

Citation

Derived from miluki/enronhop. Original pipeline and generation methodology: see the source project.