judge dict | orchestration dict |
|---|---|
{
"restarts": 15,
"verdicts": 1779,
"keep": 128,
"discard": 623,
"hold": 1028,
"errors_counter_sum": 6153,
"distinct_items": 374,
"items_judged_more_than_once": 24,
"of_those_with_differing_verdicts": 22,
"hold_share_pct": [
57.8,
55.5,
60.1
],
"discard_share_pct": [
35,
32.8... | {
"rounds": 11057,
"asks": 19141,
"asks_by_model": {
"ask_transformer": 6399,
"ask_diffusion": 6390,
"ask_rwkv": 6352
},
"asks_forced": 14526,
"forced_share_pct": [
75.9,
75.3,
76.5
]
} |
IRIS judge and orchestration counts
Counts from the judge and orchestration logs of our own pipeline: verdict mix, how often a judge model repeats itself on the same item, and how often the orchestrating model had to be forced to make a call. Counts only. The logs themselves are not published.
Files: summary.json (every number with its 95% interval) and judge_counts.py (the script that produced it from the logs). Context and key: https://indahl.ai/iris .
Terms
Free to download, read and cite with attribution (Heuremen LLC, with a link to indahl.ai). All rights reserved. Text and data mining, and use of this data to train, fine-tune or build AI models, are not permitted. Names of authors appear as published in their own papers. To have a name or a study removed: https://indahl.ai/contact . Policy: https://indahl.ai/terms#ai-training Reference HLC-A4765162.
Published by Heuremen LLC. Source: https://indahl.ai . We found it.
- Downloads last month
- 17