Qwen3-Embedding-0.6B LRAT LR 2e-6 Full Epoch
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".
Training
The model was initialized from the original Qwen3-Embedding-0.6B checkpoint and trained for one complete epoch, comprising 11,764 optimizer steps. Training used only the official LRAT trajectory-derived training pairs. The training split contains 94,113 rows and is query-disjoint from the fixed 1,500-query development set.
The experiment retains the official LRAT example weights and uses a learning rate of 2e-6. The training process was resumed from a complete checkpoint after a server restart; model, optimizer, scheduler, sampler, and random-number-generator states were restored before continuing to the predeclared final step.
Development result
On the fixed query-disjoint development set, the final checkpoint obtained Recall@1 0.630000, Recall@5 0.905333, Recall@10 0.974667, and MRR 0.752016.
This is the highest MRR point estimate among the team's query-disjoint full-epoch experiments. However, paired-bootstrap confidence intervals against the prior full-epoch candidates cross zero, so this checkpoint is submitted as an exploratory A-leaderboard candidate rather than a claimed statistically robust improvement.
The locked test split was not used for selection or evaluation.
Inference
Use last-token pooling followed by L2 normalization. The output embedding dimension is 1024.
Compliance
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.
