Qwen3 Embedding 0.6B LRAT Low-LR Continuation v1

This checkpoint continues training from the team's broad-g16-v2 retriever on the same 80,517 organizer-provided LRAT trajectory rows. The repository keeps its historical front-success-v1 experiment name, but a post-training audit confirmed that its prepared stage weights were not consumed by the cross-device loss. The effective method is therefore a conservative low-LR continuation, not stage-weighted training.

Training configuration:

• group size 16, learning rate 1.5e-7, warmup ratio 0.04, and 0.5 epoch;
• continue from broad-g16-v2, whose lineage starts from Qwen/Qwen3-Embedding-0.6B.

The input file contained a prepared 1.15 multiplier for first-30% trajectory rows. However, training used cross-device negatives with LRAT_CROSS_DEVICE_REWEIGHT left at its default 0, so these weights did not enter the effective loss. This distinction is recorded for reproducibility.

Only organizer-provided trajectories and corpus are used. No A-board query, qrel, per-query evaluation failure, external corpus, external-model label, or weights from another pretrained checkpoint enter this model.

The checkpoint was selected by aggregate fixed-seed evaluation with the released BrowseComp-Plus Agent and Judge workflow using Qwen3.5-4B. Local scores are used only for checkpoint selection and are not claimed as official results.