Claw-Eval General (English) pass^3
Agents · 2026-04-07
Claw-Eval's General split holds service-orchestration agent tasks (communication, finance, operations, productivity) graded on completion, robustness and safety from full trajectory evidence. This metric keeps the split's English-language tasks and credits a task only when all three independent trials meet the 0.75 pass threshold. Release v1.1.0's General split has 161 tasks (101 English, 60 Chinese), but English-only reports do not always use that roster or the default grader, so no item count is asserted.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-9B | 51.9 |
| LFM3-1.7B | 51.4 |
| Qwen3.5-4B | 43.0 |
| LFM2.5-2.6B | 37.5 |