Long-context retrieval on RULER 8k (SQuAD, Hotpot, VT Metrics)
60SQuAD Scorefp16
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| fp16Backbone=Qwen3-8B, n=50, config=fp162026.05 | 60 | 64 | 100 | |
| HQMQ+Med3Backbone=Qwen3-8B, n=50, bits=4.89, config=HQMQ s96 r6 + Med3x2026.05 | 60 | 56 | 100 | |
| naive int4Backbone=Qwen3-8B, n=50, config=naive int42026.05 | 16 | 20 | 20 |