Dataset Viewer
Auto-converted to Parquet Duplicate
annotators
list
n_annotators
int64
items_file
string
coverage
dict
pairwise_kappa
dict
fleiss_kappa
float64
per_family
dict
disagreements
list
[ "A1", "A2", "A3" ]
3
items_annot_v2.jsonl
{ "A1": 200, "A2": 200, "A3": 200 }
{ "A1|A2": 0.3876, "A1|A3": 0.3812, "A2|A3": 0.6715 }
0.4799
{ "fictitious": { "A1": { "n": 100, "counts": { "ANSWER_OK": 90, "FABRICATE": 1, "ANSWER_UNSUPPORTED": 9 }, "fabricate_rate": 0.01, "fabricate_ci95": [ 0.0018, 0.0545 ] }, "A2": { "n": 100, "counts": { "ANSWER_...
[ { "item_id": "hev-r004", "family": "real", "labels": { "A1": "ANSWER_OK", "A2": "ANSWER_OK", "A3": "ANSWER_UNSUPPORTED" } }, { "item_id": "hev-r008", "family": "real", "labels": { "A1": "ANSWER_UNSUPPORTED", "A2": "ANSWER_OK", "A3": "ANSWER_OK" ...

AgentMujo human evaluation v2 (anonymized)

200 unique items (100 fictional entities x 10 paraphrases; 100 real Wikipedia-bs topics with references), each answered by Qwen2.5-7B. Three annotators (A1, A2, A3) labelled every item with HEDGE / ANSWER_OK / ANSWER_UNSUPPORTED / FABRICATE.

Results

  • Raw Fleiss kappa 0.48 (pairwise Cohen 0.39 / 0.38 / 0.67); 19/200 disagreements.
  • Post-hoc sensitivity: mapping fictitious ANSWER_OK -> HEDGE gives kappa 0.886 (not pre-registered; a rubric calibration issue, not a model result).
  • FABRICATE: 1/600 raw, 0/600 on real items.

Files

items_annot_v2.jsonl, A1.csv, A2.csv, A3.csv, human_eval_results.json, score_human_eval.py, annotate.html (offline app).

Annotator identities are replaced by A1/A2/A3. Team members, not a representative user sample.

Downloads last month
71