Safety & factuality
HHEM Hallucination (Vectara)
Published by Vectara. Source page ↗ · Data we read ↗
Vectara sells a hallucination detector and runs the board with its own HHEM model as judge.
Facts
| Fact | Value | Where it comes from |
|---|---|---|
| Who runs the evaluation | a model vendor | curated by aiqbench, reviewed 2026-10-09 |
| Test set exposure | public | curated by aiqbench, reviewed 2026-10-09 |
| Data licence | Apache-2.0 | the board's attribution record |
| Newest published result | 3 days ago | the ingested rows |
| Last pulled by aiqbench | 9 hours ago · every 12 h | the sync log |
| Models on this axis | 109 | the ingested rows |
| Rows with a published interval | none published | the ingested rows |
| Headroom at the top | Best 98.2%, 1.8% below the ceiling; top 10 within 3.3%. | computed from the ingested rows |
| Axes it tracks closely | None measured yet: 7 pairs are under the shared-model floor. | Spearman ρ over shared models (see Clusters) |
Where it appears
Weighted in the built-in templates: All sources ×1.
View this axis alone in the console →Attribution
Source: Vectara Hallucination Leaderboard (vectara/hallucination-leaderboard) — https://github.com/vectara/hallucination-leaderboard. Licensed under Apache 2.0. Modified by aiqbench: we parse the ranking table embedded in the repo README between the <!-- LEADERBOARD_START --> and <!-- LEADERBOARD_END --> markers, and store factual-consistency-rate (100 − hallucination rate) as the primary score so "higher is better" holds across every board in the app. The raw hallucination rate, answer rate, and average summary length are preserved on each row in metadata.