This is a full-parameter dense retriever checkpoint based on Qwen/Qwen3-Embedding-0.6B for the CCIR Cup task "基于Agent交互轨迹的检索模型优化".

The model was trained for one complete epoch from the original Qwen3-Embedding-0.6B checkpoint. Training used only the official LRAT trajectory-derived training pairs. The training split contains 94,113 rows and is query-disjoint from the 1,500-query development set.

The only experimental change from the raw query-disjoint one-epoch run was trajectory search-order weighting. For rows with a stable mapping to the original trajectory, examples associated with later search events received weights of 1.1 or 1.2; other rows retained weight 1.0. A terminal audit verified that only reweight_rate changed. The resulting total training weight was 1.1443876568× the raw split, so this checkpoint intentionally preserves the original unnormalized experiment.

On the fixed query-disjoint development set, the checkpoint obtained Recall@1 0.6300, Recall@5 0.9060, Recall@10 0.975333, and MRR 0.749001.

Compared with the corresponding raw one-epoch checkpoint, the point estimates for Recall@1 and MRR were slightly positive, but their paired-bootstrap 95% confidence intervals crossed zero and Recall@5 decreased slightly. This checkpoint is therefore an exploratory A-leaderboard candidate rather than a claimed improvement over the team's best submitted model.

The locked test split was not used for selection or evaluation.

Use last-token pooling followed by L2 normalization. The output embedding dimension is 1024.

No external training data, external API output, teacher model, adapter, or additional pretrained model parameters were used. The repository contains a complete full-parameter checkpoint and tokenizer files at the repository root.