LLM Fine-Tuning Data Selection on DATE-LM MMLU, GSM8K, BBH (test)
61.87MMLU AccuracyBipCov (ref-aligned)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BipCov (ref-aligned)pipeline=DATE-LM, seeds=42, 1337, 2025, embeddings=Llama-3.1-8B last-token hidden-state, reference construction=prompt+label2025.02 | 61.87 | 63.53 | 66.27 | 63.89 | |
| Randompipeline=DATE-LM, seeds=42, 1337, 20252025.02 | 60 | 61.89 | 66.58 | 62.82 | |
| RDS+pipeline=DATE-LM, seeds=42, 1337, 20252025.02 | 59.58 | 62.24 | 66.8 | 62.88 |