List your company

Zero-shot Task Completion on HellaSwag 0-shot (N=500)

47.5AccuracyFull cache
39.07641.26343.4545.637Mar 20, 2026

Evaluation results

MethodLinks
Full cacheModel=Qwen2.5-1.5B2026.0347.5
KV-DirectModel=Qwen2.5-1.5B2026.0347.5
Window-64Model=Qwen2.5-1.5B2026.0347.5
Full cacheModel=Qwen3-0.6B2026.0344
KV-DirectModel=Qwen3-0.6B2026.0344
Window-64Model=Qwen3-0.6B2026.0344
Full cacheModel=DS-R1-1.5B2026.0342
KV-DirectModel=DS-R1-1.5B2026.0342
Window-64Model=DS-R1-1.5B2026.0342
Full cacheModel=Qwen2.5-0.5B2026.0341
KV-DirectModel=Qwen2.5-0.5B2026.0341
Window-64Model=Qwen2.5-0.5B2026.0341
Full cacheModel=SmolLM2-135M2026.0339.4
KV-DirectModel=SmolLM2-135M2026.0339.4
Window-64Model=SmolLM2-135M2026.0339.4