XIR 1170 — Qwen3 Embedding 0.6B — 2026-07-17 A1
Submission candidate for 基于Agent交互轨迹的检索模型优化.
- Base architecture:
Qwen3-Embedding-0.6B - Data: official LRAT competition training pairs only
- No external datasets, teacher models, model outputs, or APIs
- Pooling: last token; normalized embeddings
- Query format:
Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{query}
Training recipe
The official LRAT checkpoint was continued for 120 steps on an official-data-only set where duplicate query-positive pairs were merged, negative pools were merged by document id, and the official trajectory-derived weights were clipped and normalized. The resulting checkpoint was interpolated 50/50 with the official LRAT checkpoint.
Local validation
Query-hash-isolated validation on 2,048 official-data rows (zero query overlap):
| Model | MRR | Weighted MRR | R@1 | R@5 | R@10 |
|---|---|---|---|---|---|
| Official LRAT | 0.725830 | 0.739319 | 0.599121 | 0.893555 | 0.968750 |
| This model | 0.727131 | 0.740530 | 0.601074 | 0.897461 | 0.968750 |
This local metric is a development gate and is not the official A-board score.