Question Answering on QA Benchmark Suite Zero-shot (test)
65.3Average AccuracyWINQ
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| WINQModel Backbone=Llama-8B, Weight Quantization=2-bit, Activation Quantization=16-bit, Evaluation Protocol=zero-shot, Training Steps=150K2026.05 | 65.3 | 76.2 | 52 | 76.9 | 78.4 | 48.5 | 72.7 | 50.4 | 67.7 | |
| ParetoQModel Backbone=Llama-8B, Weight Quantization=2-bit, Activation Quantization=16-bit, Evaluation Protocol=zero-shot, Training Steps=150K2026.05 | 64.7 | 75.9 | 50.2 | 75 | 78.5 | 48.5 | 72.3 | 49.6 | 68 |