Letterweave-1.1-E2B / evaluation /private-development-aggregate.json
magnus919's picture
Keep public benchmarks at general email-labeling level
ba26936 verified
Raw History Blame Contribute Delete
2.75 kB
{
"model": "Letterweave-1.1-E2B",
"dataset": "Letterweave0.3.0",
"hardware": "NVIDIA RTX5070Ti16GB",
"quality_kind": "model_teacher known-development agreement, not human or unseen accuracy",
"training": {
"seconds": 994.4827,
"peak_allocated_mib": 9582.927734375,
"steps": 480,
"seed": 3407
},
"checkpoint": {
"decision_changes": 0,
"judgments": 1813,
"max_probability_delta": 0.0036415277772176013
},
"limits": [
"Known100guided generator design; not independent unseen test",
"Small correlated template sample; no checkpoint/threshold tuned using mail results",
"No new Jev evaluation; timing boundaries differ",
"Public comparisons contain published releases only; unmatched subsets show release1.1 standalone results."
],
"selected_checkpoint": "published Letterweave-1.1-E2B, fixed final480steps",
"comparison_scope": "Published Letterweave1.0 and1.1 only; historical published Jev results documented separately in model card",
"private100": {
"review": "original unchanged assistant/model_teacher review; known development messages, not human gold or unseen test",
"releases": {
"Letterweave-1.0-E2B": {
"acceptable": 39,
"missing_alerts": 51
},
"Letterweave-1.1-E2B": {
"acceptable": 79,
"missing_alerts": 9
}
}
},
"families": {
"meeting-recap": {
"Letterweave-1.0-E2B": {
"n": 9,
"acceptable": 0
},
"Letterweave-1.1-E2B": {
"n": 9,
"acceptable": 9
}
}
},
"synthetic_retained195": {
"Letterweave-1.1-E2B": {
"emails": 195,
"macro_f1": 0.9402415080196037,
"exact": 0.8564102564102564,
"exact_count": 167,
"per_label": {
"correspondence": {
"tp": 52,
"tn": 141,
"fp": 2
},
"transactions/records": {
"tn": 155,
"tp": 34,
"fn": 5,
"fp": 1
},
"alerts/updates": {
"tn": 103,
"tp": 85,
"fp": 6,
"fn": 1
},
"newsletters/publications": {
"tn": 180,
"tp": 12,
"fn": 3
},
"promotions/outreach": {
"tn": 168,
"tp": 25,
"fn": 1,
"fp": 1
},
"community/discussions": {
"tn": 157,
"tp": 34,
"fp": 2,
"fn": 2
},
"scheduling/invitations": {
"tn": 144,
"tp": 44,
"fp": 3,
"fn": 4
}
}
}
},
"synthetic_fresh64": {
"Letterweave-1.1-E2B": {
"macro_f1": 0.9222010425123842,
"exact_set_agreement": 0.84375
}
}
}