Image Classification on CIFAR-10 (test) (Validity and Accuracy Advantage)
64Validity CountDeepSeek6.7B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DeepSeek6.7BSource=archbest, N=32 per-arm, Protocol=Source-guided2026.07 | 64 | 50.49 | 0.2651 | 0.2693 | |
| DeepSeek6.7BSource=highsource, N=32 per-arm, Protocol=Source-guided2026.07 | 63 | 42.13 | 0.1859 | 0.1582 | |
| DeepSeek6.7BSource=highsource, N=32 per-arm, Protocol=Non-source2026.07 | 62 | 23.54 | — | — | |
| DeepSeek6.7BSource=archbest, N=32 per-arm, Protocol=Non-source2026.07 | 60 | 23.98 | — | — | |
| Olympic7BSource=archbest, N=32 per-arm, Protocol=Source-guided2026.07 | 58 | 48.06 | -0.0165 | 0.2536 | |
| Olympic7BSource=highsource, N=32 per-arm, Protocol=Source-guided2026.07 | 56 | 40.33 | -0.0843 | 0.1699 | |
| Olympic7BSource=archbest, N=32 per-arm, Protocol=Non-source2026.07 | 49 | 49.71 | — | — | |
| Qwen2.5-7BSource=highsource, N=32 per-arm, Protocol=Source-guided2026.07 | 45 | 48.14 | 0.2428 | 0.1701 | |
| Olympic7BSource=highsource, N=32 per-arm, Protocol=Non-source2026.07 | 42 | 48.76 | — | — | |
| Qwen2.5-7BSource=archbest, N=32 per-arm, Protocol=Source-guided2026.07 | 36 | 47.57 | 0.2601 | 0.23 | |
| Qwen2.5-7BSource=archbest, N=32 per-arm, Protocol=Non-source2026.07 | 30 | 21.56 | — | — | |
| Qwen2.5-7BSource=highsource, N=32 per-arm, Protocol=Non-source2026.07 | 26 | 23.86 | — | — |