Language Model Evaluation on Quality, Factuality, and Safety Evaluation Suite (test)
86.3Generation Quality ScoreSelf-Improving Pretraining
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Self-Improving PretrainingModel Scale=1.4B, Training Dataset=SlimPajama, Pretraining Objective=Pretraining for Quality2026.01 | 86.3 | 50.8 | 87.9 | 43.6 | 84.9 | |
| Self-Improving PretrainingModel Scale=1.4B, Training Dataset=SlimPajama, Pretraining Objective=Pretraining for Factuality2026.01 | 84 | 50.5 | 81.4 | 57.6 | 85.1 | |
| Self-Improving PretrainingModel Scale=1.4B, Training Dataset=RedPajama, Pretraining Objective=Pretraining for Safety2026.01 | 73.6 | 49.1 | 73.9 | 38 | 91.1 | |
| Llama-3.1 8B BaseModel Scale=8B2026.01 | 66.1 | 63.1 | 77.1 | 26.3 | 71 | |
| Llama Pretrain BaselineModel Scale=1.4B, Training Dataset=RedPajama, Pretraining Objective=Pretraining for Safety2026.01 | 54.5 | 47.9 | 57.1 | 40.8 | 75.5 | |
| Llama BaseModel Scale=1.4B2026.01 | 50 | 47.6 | 50.1 | 42.3 | 76.9 | |
| Llama Pretrain BaselineModel Scale=1.4B, Training Dataset=SlimPajama2026.01 | 49 | 46.8 | 49.4 | 44 | 76.9 |