Steve0927/qwen3-embedding-0.6b-lrat-global256-utility-weighted-v1
This is a Qwen3-Embedding-0.6B dense retriever checkpoint for XIR competition 1170.
Method
Cross-Device Trajectory Utility Weighting with conservative row filtering
The released eight-rank LRAT recipe has a global batch of 256. On four H200 GPUs we preserve that retrieval geometry with batch 64 per rank and scale the pre-clipping loss by 4/8, matching the released DDP gradient normalization. The weighted variants also preserve each official trajectory row's reweight_rate after cross-device embedding gathering; the inherited decoder-only path otherwise discards those weights.
Training uses 80,517 rows deterministically filtered from the official LRAT training-pairs JSONL. The final continuation is 0.5 epoch at 5e-7 from the calibrated reproduction.
Data Boundary
- Original model weights:
Qwen/Qwen3-Embedding-0.6Bonly. - Training rows and all supervision come from the organizer-provided LRAT trajectories and offline corpus.
- No generated queries, answers, passages, or external evidence.
- No A/B qrels, query-level failures, or leaderboard cases enter training.
- Evaluation results are used only as aggregate checkpoint gates.
Evaluation
Full 830-query local A-board evaluation: total=38.0361, recall=44.1349, success=23.8554, average_steps=22.9000, step_score=54.2000.
The local evaluation uses the released BrowseComp-Plus A-board pipeline with the fixed Qwen3.5-4B Agent/Judge, a 50-search budget, and the organizer's score formula.
Intended Use
Use this model as the dense retriever in the fixed-corpus LRAT/BrowseComp-Plus agent-search setting.