Search Ranking on Unseen search data (test)
11.8M vs V2 ScoreV1
Evaluation Results
| Method | Links | |
|---|---|---|
| V1Key Innovations=Mean-Pooling + DCN + MOE, Seq Len=–, LR=5e−5, Status=Baseline2026.03 | 11.8 | |
| V3.5Key Innovations=V3.4 + Four-phase LR optimization (Table 2), Seq Len=200, LR=Adaptive, Status=New Best2026.03 | 8.3 | |
| V3.4Key Innovations=V3.3 + Temporal Embeddings, Seq Len=200, LR=1e−5, Status=Success2026.03 | 6.3 | |
| V3.3Key Innovations=V3.2 + Slot Type Embeddings, Seq Len=200, LR=1e−5, Status=Success2026.03 | 6.1 | |
| V3.1Key Innovations=Transformer + Unified Sequence (Figure 3b), Seq Len=200, LR=5e−5, Status=Failed2026.03 | 5 | |
| V3.0Key Innovations=V2 + Positional Encoding + Attention Pooling, Seq Len=40, LR=5e−5, Status=Failed2026.03 | 4.7 | |
| V3.2Key Innovations=V3.1 + Reduced LR, Seq Len=200, LR=1e−5, Status=Success2026.03 | 4.1 | |
| V2Key Innovations=Transformer + Separate Sequences (Figure 3a), Seq Len=40, LR=5e−5, Status=New Baseline2026.03 | 0 |