The methodology described here: https://github.com/CerebrasResearch/reap
Applied to Qwen/Qwen3-30B-A3B-Instruct-2507 with 50% compression, with super experts singleton.
Base dataset for pruning: cnmoro/PTBR-REAP-Pruning
The methodology described here: https://github.com/CerebrasResearch/reap
Applied to Qwen/Qwen3-30B-A3B-Instruct-2507 with 50% compression, with super experts singleton.
Base dataset for pruning: cnmoro/PTBR-REAP-Pruning