Based on the official LRAT Qwen3-Embedding-0.6B checkpoint.

Training method: 90-step low-learning-rate continuation on the deduplicated sqrt-weight dataset, followed by BF16 interpolation with the current best checkpoint using candidate weight 0.10 and base weight 0.90.

Local validation on 2048 held-out rows: MRR 0.727729, weighted MRR 0.741520, Recall@1 0.602051, Recall@5 0.895020.