Training Data Composition Estimation on GPT-2 mixtures (controlled sandbox)
87.53Overlap AccuracyLLMSurgeon
Evaluation results
| Method | Links | |
|---|---|---|
| LLMSurgeonGPT-2 Model=gpt2_web_heavy2026.05 | 87.53 | |
| LLMSurgeonGPT-2 Model=gpt2_balanced2026.05 | 75.62 | |
| LLMSurgeonGPT-2 Model=gpt2_book_heavy2026.05 | 50.15 |