LRAT Qwen3 Embedding 0.6B — M07
This is the M07 retrieval model prepared for the CCIR Cup / LRAT evaluation. It is a full-parameter fine-tune of Qwen/Qwen3-Embedding-0.6B.
Training
- Data: 94,113 official LRAT training rows after deterministic normalized-query separation
- Training length: one epoch, 11,764 optimizer steps
- Per-device batch size: 1 on two GPUs
- Gradient accumulation: 4
- Learning rate: 1e-6
- External training data, teacher models, LoRA, and adapters: not used
The separated development set contains 1,500 normalized-query groups with zero query overlap with training.
Development metrics
| Metric | Value |
|---|---|
| Recall@1 | 0.626667 |
| Recall@5 | 0.907333 |
| Recall@10 | 0.974667 |
| MRR | 0.747510 |
These are candidate-set development metrics, not official leaderboard results.
Integrity
model.safetensors SHA-256:
c937d61d57840467c5b8867559cc645704313e97ace7dafdeeece52652986a2b