Long-context understanding on LongBench-Pro English
37.25AccuracyUNIQUE
Evaluation results
| Method | Links | |
|---|---|---|
| UNIQUEBudget=512, Fine-tuning protocol=LoRA fine-tuned on LongAlign-10k, Base Model=Ministral-3-8B-Instruct-25122026.05 | 37.25 | |
| DSABudget=512, Fine-tuning protocol=LoRA fine-tuned on LongAlign-10k, Base Model=Ministral-3-8B-Instruct-25122026.05 | 36.79 | |
| Full attentionFine-tuning protocol=LoRA fine-tuned on LongAlign-10k, Base Model=Ministral-3-8B-Instruct-2512, Budget=Full attention2026.05 | 36.55 | |
| InfLLM-v2Budget=512, Fine-tuning protocol=LoRA fine-tuned on LongAlign-10k, Base Model=Ministral-3-8B-Instruct-25122026.05 | 36.11 |