marcusguhao/IndexMem-Qwen3-8B
Updated
Learned KV-cache eviction gates for Qwen3, from "The Gating Mechanism of KV Cache Eviction".
Note Retention gate for a frozen Qwen3-8B. RULER 8K 90.14 at 75% compression.
Note Retention gate for a frozen Qwen3-4B. RULER 8K 83.89 at 75% compression.