Image Classification Code Generation on MNIST, CelebA, SVHN, ImageNette, CIFAR-10, CIFAR-100 (val)
76Valid RateIterative
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| IterativeLLM=DS-6.7B, Fine-tuned=No, Cycles=2000§, Total Gen.=2,000, Datasets=32026.05 | 76 | — | — | 71.5 | — | — | 200 | |
| Delta-Code GenerationLLM=DS-7B, Fine-tuned=Yes, Cycles=22, Total Gen.=1,100, Datasets=62026.05 | 75.3 | 72.677 | 65.8 | 99.5 | 75.7 | 72.778 | 30.4 | |
| Delta-Code GenerationLLM=Qwen-7B, Fine-tuned=Yes, Cycles=22, Total Gen.=1,100, Datasets=62026.05 | 72.1 | 69.474 | 64.6 | 99.5 | 74.4 | 71.377 | 31.4 | |
| Delta-Code GenerationLLM=Mistral-7B, Fine-tuned=Yes, Cycles=22, Total Gen.=1,100, Datasets=62026.05 | 66.6 | 63.869 | 66.1 | 99.5 | 75.9 | 72.678 | 49.5 | |
| Full Gen.LLM=DS-7B, Fine-tuned=Yes, Cycles=22, Total Gen.=1,100, Datasets=12026.05 | 50.6 | 45.056 | 42.3 | 64 | 51.1 | — | 200 |