Update README.md

README.md

@@ -89,15 +89,9 @@ The model was trained on 80,000+ queries from 11 benchmarks:
 
 | Domain | Datasets |
 |--------|----------|
-| Financial | FinReport, FinSlides, FinQA, ConvFinQA |
+| Financial | FinReport, FinSlides, FinQA, ConvFinQA, TAT-DQA |
 | Scientific | ArxivQA, SciQAG |
-| General | Wiki-SS, MP-DocVQA, DUDE, VQAnBD, TAT-DQA |
-
-### Training Procedure
-
-1. **Oracle Label Generation**: Run all retrieval pipelines on training queries to collect nDCG@5 and latency metrics
-2. **Reward Computation**: `r(q, i) = (1 - λ) · nDCG(q, i) + λ · (1 - NormalizedLatency(q, i))`
-3. **Soft Label Training**: Train with weighted KL divergence loss using reward scores as soft labels
+| General | Wiki-SS, MP-DocVQA, DUDE, VQAnBD,  |
 
 ### Hyperparameters
 
@@ -112,14 +106,6 @@ The model was trained on 80,000+ queries from 11 benchmarks:
 | Precision | bfloat16 |
 | λ (trade-off) | 0.0 (accuracy-focused) |
 
-## Performance
-
-### Latency
-
-| Component | Time |
-|-----------|------|
-| Router Inference | ~15ms |
-
 ## Intended Use
 
 IRouterLM is designed for:
@@ -132,7 +118,6 @@ IRouterLM is designed for:
 
 - Trained on English queries only
 - Optimized for document retrieval tasks (financial, scientific, general domains)
-- Requires the corresponding retrieval pipelines to be available
 
 ## License