Cross-split leakage audit
Review overlapping training and evaluation text before trusting model scores. Each finding points to the original one-based JSONL line numbers.
Public walkthrough · All excerpts are original synthetic examples. Source & CLI guide ↗
Split coverage
Each affected row is counted once per split. Percentages use all submitted rows, including skipped records.
| Split | Submitted | Eligible | Affected / all | Skipped by reason |
|---|---|---|---|---|
| train | 8 | 8 | 4 (50.0%) | — |
| validation | 4 | 4 | 2 (50.0%) | — |
| test | 7 | 5 | 3 (42.9%) | number or identifier: 1, punctuation or symbols: 1 |
Where overlap occurs
| Split pair | Normalized exact | Character Jaccard | One short edit |
|---|---|---|---|
| test ↔ train | 1 | 1 | 1 |
| test ↔ validation | 1 | 0 | 0 |
| train ↔ validation | 2 | 0 | 0 |
Evidence register
One representative per matching text-group / split pair, up to 100. 6 of 6 evidence groups included. Where available the representative has conflicting labels.
exacttrain:6 ↔ validation:3 · 1 row pair(s) · 0 label conflict(s)
Identical after normalization (10 characters).
wind speed
wind speed
exacttrain:6 ↔ test:3 · 1 row pair(s) · 0 label conflict(s)
Identical after normalization (10 characters).
wind speed
wind speed
exactvalidation:3 ↔ test:3 · 1 row pair(s) · 0 label conflict(s)
Identical after normalization (10 characters).
wind speed
wind speed
exacttrain:1 ↔ validation:1 · 1 row pair(s) · 1 label conflict(s)
Identical after normalization (12 characters).
请检查a区的温度传感器。
请检查a区的温度传感器。
char_jaccardtest:2 ↔ train:3 · 1 row pair(s) · 0 label conflict(s)
66 shared character shingles out of 68 in their union. Jaccard 0.971.
at sunrise record the compass heading from the eastern edge of the old pier!
at sunrise record the compass heading from the eastern edge of the old pier.
short_edittrain:2 ↔ test:1 · 1 row pair(s) · 0 label conflict(s)
One substitution at normalized character offset 5 (zero-based).
关闭仓库照明灯
关闭仓库照眀灯
Run details
Includes normalized text excerpts. Treat this report as source data.
Character shingle width 3. Minimum Jaccard 0.74. Short one-edit rule on for normalized lengths 6–18. Search visited 18 of 2,000,000 allowed posting entries and verified 2 candidates.
A close lexical match can carry a different meaning. Inspect the source rows before merging or removing examples. Label disagreements use only pairs with known labels on both sides (6).