Run History
Every saved run, newest first. Extended runs carry the full metric taxonomy and link through to the Analysis view.
2026-09-08 05:56:40
MMLU Β· extended
2850 items
JSON
task quality β
consistency β
robustness β
api performance β
token efficiency β
economics β
reliability β
1 stage not run
top_p=1.0
max_tokens=384
top_k=1
repetition_penalty=1.0
seed=42
benchmark=mmlu
subjects=57 subject(s), 50 q each
TP1Β·PP1Β·DP1Β·SP1Β·CP1Β·EP1
(provider-chosen)
| # | Model | Accuracy (%) | Correct | Total | Avg time (s) | Composite | $ / correct | Provider |
|---|---|---|---|---|---|---|---|---|
| 1 | Ministral-8B | 79.4 | 2262 | 2850 | 1.406 | 0.798 | $0.000051 | Mistral |
| 2 | Llama-3.2-3B | 54.7 | 1560 | 2850 | 0.693 | 0.707 | $0.000083 | Cloudflare |
| 3 | Gemma-3-4B | 61.2 | 1743 | 2850 | 2.21 | 0.656 | $0.000046 | DeepInfra |
β Models in this run were scored over different component sets, so their composites use different effective weights and are not directly comparable.
1 stage have no data
Probabilistic quality, consistency and robustness need extra passes over
the dataset. Run them now on a sample to fill those sections.
2026-09-07 23:52:20
LAMBADA Β· extended
1000 items
JSON
task quality β
context β
api performance β
token efficiency β
economics β
reliability β
tokenization β
3 stages not run
top_p=1.0
max_tokens=8
few_shot=2
TP1Β·PP1Β·DP1Β·SP1Β·CP1Β·EP1
(provider-chosen)
| # | Model | Accuracy (%) | Correct | Total | Avg time (s) | Composite | $ / correct | Provider |
|---|---|---|---|---|---|---|---|---|
| 1 | Ministral-8B | 39.5 | 395 | 1000 | 0.518 | 0.517 | $0.000096 | Mistral |
| 2 | Llama-3.2-3B | 22.4 | 224 | 1000 | 0.349 | 0.446 | $0.000101 | Cloudflare |
| 3 | Gemma-3-4B | 18.0 | 180 | 1000 | 1.812 | 0.365 | $0.000107 | DeepInfra |
Composite weighted over quality, efficiency, reliability β calibration and robustness not measured in this run, so the weight is redistributed over the rest rather than scored zero.
3 stages have no data
Probabilistic quality, consistency and robustness need extra passes over
the dataset. Run them now on a sample to fill those sections.