Update README.md

README.md

@@ -45,9 +45,7 @@ widget:
 
 # Avey-B Urdu 25M
 
-**A compact, fast Urdu encoder trained from scratch for masked-language modeling.**
-
- **24.87M parameters** ·  **408K tokens/s on RTX 4060** ·  **32K Urdu WordPiece vocabulary**
+A compact, fast Urdu encoder trained from scratch for masked-language modeling.
 
 ![Quality and efficiency overview](assets/benchmark_overview.png)
 
@@ -57,7 +55,7 @@ widget:
 - **81.27 WikiANN Urdu entity F1** with only 24.87M parameters.
 - **1.71 ms** encoder latency at batch 1 / length 128 and **148 MiB** peak VRAM at batch 16 / length 512.
 - **14.29 held-out MLM perplexity**, **50.46% top-1**, and **70.39% top-5** masked-token accuracy.
-- Compact Urdu tokenization: **4.19 characters/token** on held-out data.
+- Compact Urdu tokenization: **4.19 characters/token** on held out data.
 
 ##  Quick start
 
@@ -101,18 +99,6 @@ print(hidden.shape)  # [batch, tokens, 384]
 | HPLT Urdu BERT | 124.36M | 83.93% ± 1.27 | 90.92 |
 | XLM-R base | 278.04M | 78.80% ± 0.52 | 84.89 |
 
-Sentiment uses 4K train / 1K validation / 1K test examples for two epochs; the test result is mean ± sample standard deviation over three seeds after validation learning-rate selection. The source is a machine-translated Urdu IMDb dataset. NER uses 4K training examples and the full 1K validation / 1K test Urdu WikiANN splits for three epochs; it reports validation-selected entity F1 for one seed. See [`benchmark_results.json`](benchmark_results.json) for machine-readable results.
-
-### RTX 4060 BF16 inference
-
-| Model | B1/L128 latency | B16/L512 throughput | B16/L512 peak VRAM |
-|---|---:|---:|---:|
-| **Avey-B Urdu** | **1.71 ms** | **408,329 tok/s** | **148 MiB** |
-| UrNova | 4.98 ms | 84,695 tok/s | 345 MiB |
-| HPLT Urdu BERT | 4.95 ms | 40,400 tok/s | 1,821 MiB |
-| XLM-R base | 2.18 ms | 136,274 tok/s | 678 MiB |
-
-Eager encoder inference with synthetic full-length token tensors, 10 warmup and 30 measured iterations. Hardware and software choices can materially change these numbers.
 
 ## Training
 
@@ -129,7 +115,6 @@ Eager encoder inference with synthetic full-length token tensors, 10 warmup and
 | Hardware/runtime | NVIDIA RTX 4060 8GB, approximately 20 minutes |
 | Final train loss | 2.52 |
 
-The architecture has 12 alternating static/dynamic Avey layers, 384-dimensional embeddings, 128-token chunks, and `k=3` retrieval/compression context. The checkpoint was trained at sequence length 512; longer contexts were not evaluated here.
 
 ## Intended use and limitations
 
@@ -137,7 +122,7 @@ Use this checkpoint for Urdu classification, token classification, representatio
 
 HPLT packages its dataset under CC0 but does not own the underlying web text. The Apache-2.0 model license does not grant rights to third-party source documents that may be reproduced by the model.
 
-## 📄 Architecture and citation
+## Architecture and citation
 
 Based on [Avey-B](https://arxiv.org/abs/2602.15814), an attention-free bidirectional encoder architecture.