9.7 GB
artur7236's picture
stack + multi-task SFT (mega 14.9k, lr=3e-7, 2500 steps) β€” v28 5-shard MIN=0.375 MEAN=0.583 worst-axis, KL=0.152, code_bench 0.62 on 4/5 shards
ade5749 verified