Baobao Algorithm Notes
Nov 5, 2026 · Artificial Intelligence
15 Pitfalls in Training JEV Decision Models: From Pointer Heads to Data Leakage
The author shares 15 practical pitfalls encountered while training JEV decision models on Qwen3.5, covering pointer head design, bidirectional attention, data loading, label-sorted collapse, shortcut learning, catastrophic forgetting, joint classification-extraction trade-offs, extraction threshold calibration, and engineering bugs like padding leaks and NCCL deadlocks.
GGUFJEV decision modelQwen3.5
0 likes · 43 min read
