Publish Avey-B Urdu 25M with benchmarks and model card

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/benchmark_overview.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Avey Inc.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

MANIFEST.json

@@ -0,0 +1,56 @@
+{
+  "generated_at": "2026-09-10T09:59:10.314711+00:00",
+  "lang": "urd_Arab",
+  "map_url": "https://data.hplt-project.org/three/sorted/urd_Arab.map",
+  "bins_requested": [
+    10,
+    9
+  ],
+  "prob_threshold": 0.8,
+  "min_chars": 20,
+  "val_fraction": 0.005,
+  "max_docs_per_shard": 50000,
+  "max_chars_per_shard": 500000000,
+  "max_shards": null,
+  "stream_download": true,
+  "shards": [
+    {
+      "url": "https://data.hplt-project.org/three/sorted/urd_Arab/10_1.jsonl.zst",
+      "bin": 10,
+      "raw_docs": 2162,
+      "dropped_low_prob": 0,
+      "dropped_empty": 0,
+      "dropped_exact_dup": 0,
+      "pii_spans_redacted": 24,
+      "kept_train": 2151,
+      "kept_val": 11,
+      "chars_kept": 41055559
+    },
+    {
+      "url": "https://data.hplt-project.org/three/sorted/urd_Arab/9_1.jsonl.zst",
+      "bin": 9,
+      "raw_docs": 29290,
+      "dropped_low_prob": 3,
+      "dropped_empty": 0,
+      "dropped_exact_dup": 1,
+      "pii_spans_redacted": 729,
+      "kept_train": 29146,
+      "kept_val": 140,
+      "chars_kept": 500109810
+    }
+  ],
+  "totals": {
+    "raw_docs": 31452,
+    "kept_train": 31297,
+    "kept_val": 151,
+    "dropped_low_prob": 3,
+    "dropped_empty": 0,
+    "dropped_exact_dup": 1,
+    "pii_spans_redacted": 753,
+    "chars_kept": 541165369
+  },
+  "consolidated_counts": {
+    "train": 31297,
+    "val": 151
+  }
+}

README.md

@@ -0,0 +1,153 @@
+---
+license: apache-2.0
+language:
+- ur
+library_name: transformers
+pipeline_tag: fill-mask
+tags:
+- avey-b
+- urdu
+- masked-language-modeling
+- encoder
+- custom_code
+datasets:
+- HPLT/HPLT3.0
+model-index:
+- name: Avey-B Urdu 25M
+  results:
+  - task:
+      type: text-classification
+      name: Urdu sentiment classification
+    dataset:
+      type: mirfan899/imdb_urdu_reviews
+      name: IMDb Urdu Reviews
+      split: test
+    metrics:
+    - type: accuracy
+      value: 0.785333
+      name: Accuracy
+  - task:
+      type: token-classification
+      name: Urdu named-entity recognition
+    dataset:
+      type: unimelb-nlp/wikiann
+      name: WikiANN Urdu
+      config: ur
+      split: test
+    metrics:
+    - type: f1
+      value: 0.812677
+      name: Entity F1
+widget:
+- text: "پاکستان کا دارالحکومت [MASK] ہے۔"
+- text: "اردو ایک خوبصورت [MASK] ہے۔"
+---
+
+# 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**
+
+![Quality and efficiency overview](assets/benchmark_overview.png)
+
+## ✨ Highlights
+
+- **78.53% ± 0.51 sentiment accuracy**, essentially level with 278M-parameter XLM-R base in this setup.
+- **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.
+
+## 🧪 Quick start
+
+This model contains a custom Avey architecture, so `trust_remote_code=True` is required.
+
+```python
+import torch
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+repo_id = "mahwizzzz/avey-b-ur"
+tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
+model = AutoModelForMaskedLM.from_pretrained(repo_id, trust_remote_code=True).eval()
+
+text = "پاکستان کا دارالحکومت [MASK] ہے۔"
+inputs = tokenizer(text, return_tensors="pt")
+with torch.inference_mode():
+    logits = model(**inputs).logits
+
+mask_index = (inputs.input_ids[0] == tokenizer.mask_token_id).nonzero()[0, 0]
+top_ids = logits[0, mask_index].topk(5).indices
+print(tokenizer.convert_ids_to_tokens(top_ids.tolist()))
+```
+
+Extract contextual token representations with `AutoModel`:
+
+```python
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("mahwizzzz/avey-b-ur", trust_remote_code=True)
+encoder = AutoModel.from_pretrained("mahwizzzz/avey-b-ur", trust_remote_code=True)
+hidden = encoder(**tokenizer("یہ ایک اردو جملہ ہے۔", return_tensors="pt")).last_hidden_state
+print(hidden.shape)  # [batch, tokens, 384]
+```
+
+## 📊 Evaluation
+
+| Model | Parameters | Sentiment accuracy | WikiANN Urdu NER F1 |
+|---|---:|---:|---:|
+| **Avey-B Urdu** | **24.87M** | **78.53% ± 0.51** | **81.27** |
+| UrNova | 94.72M | 81.83% ± 0.42 | 87.79 |
+| 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
+
+![Training loss](assets/training_loss.png)
+
+| Item | Value |
+|---|---|
+| Corpus | HPLT 3.0 Urdu (`urd_Arab`), quality bins 10 and 9 |
+| Clean data | 31,297 train docs, 151 validation docs, 541,165,369 characters |
+| Sequence length | 512 |
+| Token batch | 16,384 tokens/step |
+| Training | 20,000 steps, BF16, fused AdamW |
+| Peak learning rate | 7e-4 |
+| 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
+
+Use this checkpoint for Urdu classification, token classification, representation extraction, or continued pretraining. It is an **encoder**, not a chat model or free-text generator. Benchmark coverage is limited to masked-token prediction, machine-translated movie-review sentiment, and Wikipedia-derived NER; evaluate on native, domain-specific Urdu before production use.
+
+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
+
+Based on [Avey-B](https://arxiv.org/abs/2602.15814), an attention-free bidirectional encoder architecture.
+
+```bibtex
+@inproceedings{2026aveyb,
+  title={Avey-B},
+  author={Acharya, Devang and Hammoud, Mohammad},
+  booktitle={The Fourteenth International Conference on Learning Representations},
+  year={2026}
+}
+```
+
+Training provenance is available in [`MANIFEST.json`](MANIFEST.json). Model code is included for reproducible loading.

benchmark_results.json

@@ -0,0 +1,159 @@
+{
+  "protocol": {
+    "sentiment": "4k train / 1k validation / 1k test, 2 epochs, 3 seeds",
+    "ner": "Urdu WikiANN, 4k train / 1k validation / 1k test, 3 epochs, seed 0",
+    "efficiency": "RTX 4060, eager BF16 encoder inference, 10 warmup + 30 measured iterations"
+  },
+  "models": {
+    "Avey-B Urdu": {
+      "model_id": "avey-b-urdu-25m/checkpoint_19999",
+      "sentiment_accuracy_mean": 0.7853333333333333,
+      "sentiment_accuracy_std": 0.005131601439446889,
+      "sentiment_seed_scores": [
+        0.781,
+        0.784,
+        0.791
+      ],
+      "sentiment_best_lr": 0.0005,
+      "wikiann_ner_f1": 0.8126773888363292,
+      "wikiann_best_lr": 0.0005,
+      "parameters": 24865804,
+      "checkpoint_mb": 94.86239624023438,
+      "latency_ms_b1_l128": 1.7135274251302084,
+      "tokens_per_second_b16_l512": 408328.5424109701,
+      "peak_vram_mb_b16_l512": 148.267578125
+    },
+    "UrNova": {
+      "model_id": "ProximaAI/urnova-95m",
+      "sentiment_accuracy_mean": 0.8183333333333332,
+      "sentiment_accuracy_std": 0.0041633319989322695,
+      "sentiment_seed_scores": [
+        0.817,
+        0.815,
+        0.823
+      ],
+      "sentiment_best_lr": 0.0001,
+      "wikiann_ner_f1": 0.877932024892293,
+      "wikiann_best_lr": 0.0001,
+      "parameters": 94723200,
+      "checkpoint_mb": 363.1072540283203,
+      "latency_ms_b1_l128": 4.984183247884115,
+      "tokens_per_second_b16_l512": 84695.40102035012,
+      "peak_vram_mb_b16_l512": 344.56005859375
+    },
+    "HPLT Urdu BERT": {
+      "model_id": "HPLT/hplt_bert_base_ur",
+      "sentiment_accuracy_mean": 0.8393333333333333,
+      "sentiment_accuracy_std": 0.012662279942148398,
+      "sentiment_seed_scores": [
+        0.853,
+        0.828,
+        0.837
+      ],
+      "sentiment_best_lr": 0.0001,
+      "wikiann_ner_f1": 0.9091792132102962,
+      "wikiann_best_lr": 0.0001,
+      "parameters": 124361472,
+      "checkpoint_mb": 596.7949447631836,
+      "latency_ms_b1_l128": 4.946500142415364,
+      "tokens_per_second_b16_l512": 40399.82395789245,
+      "peak_vram_mb_b16_l512": 1821.2353515625
+    },
+    "XLM-R base": {
+      "model_id": "FacebookAI/xlm-roberta-base",
+      "sentiment_accuracy_mean": 0.788,
+      "sentiment_accuracy_std": 0.005196152422706636,
+      "sentiment_seed_scores": [
+        0.791,
+        0.782,
+        0.791
+      ],
+      "sentiment_best_lr": 2e-05,
+      "wikiann_ner_f1": 0.8488652824722356,
+      "wikiann_best_lr": 2e-05,
+      "parameters": 278043648,
+      "checkpoint_mb": 1063.8882179260254,
+      "latency_ms_b1_l128": 2.1774335225423176,
+      "tokens_per_second_b16_l512": 136274.27949161586,
+      "peak_vram_mb_b16_l512": 677.99609375
+    }
+  },
+  "efficiency_protocol": {
+    "device": "NVIDIA GeForce RTX 4060",
+    "dtype": "bfloat16",
+    "mode": "eager inference",
+    "warmup": 10,
+    "iterations": 30
+  },
+  "held_out_mlm": {
+    "protocol": {
+      "validation_data": "data/processed/val",
+      "batch_size": 2,
+      "sequence_length": 512,
+      "batches": 100,
+      "mask_fraction": 0.15,
+      "seed": 2026,
+      "dtype": "bfloat16"
+    },
+    "checkpoints": {
+      "random_init": {
+        "masked_tokens": 15378,
+        "loss": 10.379354222169049,
+        "perplexity": 32188.168372746946,
+        "top1_accuracy": 6.502796202367018e-05,
+        "top5_accuracy": 0.00026011184809468074,
+        "parameters": 24865804
+      },
+      "checkpoint_999": {
+        "masked_tokens": 15378,
+        "loss": 6.6500120069940625,
+        "perplexity": 772.7936044076636,
+        "top1_accuracy": 0.1179607231109377,
+        "top5_accuracy": 0.21335674339966185,
+        "parameters": 24865804
+      },
+      "checkpoint_9999": {
+        "masked_tokens": 15378,
+        "loss": 3.1010802833695053,
+        "perplexity": 22.221944316353014,
+        "top1_accuracy": 0.4473923787228508,
+        "top5_accuracy": 0.6474183899076603,
+        "parameters": 24865804
+      },
+      "checkpoint_19999": {
+        "masked_tokens": 15378,
+        "loss": 2.6596765298653615,
+        "perplexity": 14.291665423968677,
+        "top1_accuracy": 0.5046169853036806,
+        "top5_accuracy": 0.703862660944206,
+        "parameters": 24865804
+      }
+    }
+  },
+  "tokenizer_efficiency": {
+    "documents": 151,
+    "characters": 2739705,
+    "models": {
+      "avey-b-urdu-25m/checkpoint_19999": {
+        "tokens": 654351,
+        "tokens_per_character": 0.23883994809660164,
+        "characters_per_token": 4.186904276145372
+      },
+      "ProximaAI/urnova-95m": {
+        "tokens": 667376,
+        "tokens_per_character": 0.24359410958479105,
+        "characters_per_token": 4.105189578288701
+      },
+      "HPLT/hplt_bert_base_ur": {
+        "tokens": 693627,
+        "tokens_per_character": 0.2531757981242506,
+        "characters_per_token": 3.9498246175538148
+      },
+      "FacebookAI/xlm-roberta-base": {
+        "tokens": 763456,
+        "tokens_per_character": 0.2786635787429669,
+        "characters_per_token": 3.5885565114427025
+      }
+    }
+  }
+}

config.json

@@ -0,0 +1,26 @@
+{
+  "architectures": [
+    "AveyForMaskedLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_avey.AveyConfig",
+    "AutoModel": "modeling_avey.AveyModel",
+    "AutoModelForMaskedLM": "modeling_avey.AveyForMaskedLM",
+    "AutoModelForSequenceClassification": "modeling_avey.AveyForSequenceClassification",
+    "AutoModelForTokenClassification": "modeling_avey.AveyForTokenClassification"
+  },
+  "chunk_size": 128,
+  "context_proportion": 0.5,
+  "d_embed": 384,
+  "dtype": "float32",
+  "eps": 1e-12,
+  "expansion_factor": 4,
+  "hidden_size": 384,
+  "k": 3,
+  "max_position_embeddings": 4294967296,
+  "model_type": "avey-b",
+  "n_layers": 12,
+  "training_sequence_length": 512,
+  "transformers_version": "5.16.1",
+  "vocab_size": 32000
+}

configuration_avey.py

@@ -0,0 +1,28 @@
+from transformers import PretrainedConfig
+
+
+class AveyConfig(PretrainedConfig):
+    model_type = "avey-b"
+
+    def __init__(
+        self,
+        vocab_size: int = 50304,
+        context_len: int = 512,
+        d_embed: int = 768,
+        n_layers: int = 26,
+        chunk_size: int = 128,
+        k: int = 3,
+        eps=1e-12,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.d_embed = d_embed
+        self.n_layers = n_layers
+        self.chunk_size = chunk_size
+        self.k = k
+        self.eps = eps
+
+        # for compatibility with the eval lib
+        self.max_position_embeddings = context_len
+        self.hidden_size = d_embed
+        super().__init__(**kwargs)

fertility_check.json

@@ -0,0 +1,3 @@
+{
+  "avey-b-urdu-tokenizer": 0.23883994809660164
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2b420e51caaf7a43267a7ca17a5532327dd96382b0d9294b8d03edeec0b26bb
+size 99470432

modeling_avey.py

@@ -0,0 +1,307 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers import PreTrainedModel
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+
+from .configuration_avey import AveyConfig
+
+
+class StaticLayer(nn.Module):
+    def __init__(self, config: AveyConfig):
+        super().__init__()
+        self.norm = nn.RMSNorm(config.d_embed, eps=config.eps)
+        self.enricher = nn.Linear(config.d_embed, config.d_embed * 4)
+
+        proj_size = config.chunk_size
+        self.spatial_proj = nn.Parameter(torch.empty(proj_size, proj_size))
+        nn.init.xavier_normal_(self.spatial_proj)
+
+        self.fuser = nn.Linear(int(config.d_embed * 3), config.d_embed)
+        self.alpha = nn.Parameter(torch.tensor(1.0))
+
+    def forward(self, x, attention_mask=None):
+        _, T, _ = x.shape
+        res = x
+        x = self.norm(x)
+        x = self.enricher(x)
+        x = F.gelu(x)
+        if attention_mask is not None:
+            x = x * attention_mask
+        x, bypass = x.chunk(2, dim=-1)
+
+        x, gate = x.chunk(2, dim=-1)
+        x = self.spatial_proj[:T, :T] @ x
+        x = gate * x
+
+        x = torch.cat([x, bypass], dim=-1)
+        x = self.fuser(x)
+        x = x + (self.alpha * res)
+        return x if attention_mask is None else x * attention_mask
+
+
+class DynamicLayer(nn.Module):
+    def __init__(self, config: AveyConfig):
+        super().__init__()
+        self.norm = nn.RMSNorm(config.d_embed, eps=config.eps)
+        self.enricher = nn.Linear(config.d_embed, config.d_embed * 4)
+        self.fuser = nn.Linear(config.d_embed * 3, config.d_embed)
+        self.alpha = nn.Parameter(torch.tensor(1.0))
+
+    def forward(self, x, attention_mask=None):
+        _, T, _ = x.shape
+        res = x
+        x = self.norm(x)
+        x = self.enricher(x)
+        x = F.gelu(x)
+        if attention_mask is not None:
+            x = x * attention_mask
+        x, bypass = x.chunk(2, dim=-1)
+
+        x, gate = x.chunk(2, dim=-1)
+        x_norm = F.normalize(x, p=2, dim=-1)
+        sim_scores = x_norm @ x_norm.mT
+        x = F.normalize(sim_scores, p=1, dim=-1) @ x
+        x = gate * x
+
+        x = torch.cat([x, bypass], dim=-1)
+        x = self.fuser(x)
+        x = x + (self.alpha * res)
+        return x if attention_mask is None else x * attention_mask
+
+
+class Ranker(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size = config.chunk_size
+        self.k = config.k + 1
+        self.extended_len = self.k * config.chunk_size
+        self.eps = config.eps
+        self.down_proj = nn.Parameter(torch.empty(self.chunk_size, self.extended_len))
+        nn.init.xavier_normal_(self.down_proj)
+
+    def preprocess(self, x):
+        B, T, E = x.shape
+        cs, L = self.chunk_size, self.extended_len
+
+        padded = False
+        orig_T = T
+        if T % cs != 0:
+            pad_len = cs - (T % cs)
+            pad = torch.zeros(B, pad_len, E, device=x.device, dtype=x.dtype)
+            x = torch.cat([x, pad], dim=1)
+            T += pad_len
+            padded = True
+
+        N = T // cs
+        x_chunks = x.view(B, N, cs, E)
+
+        extended = []
+        for i in range(0, N):
+            cur = x_chunks[:, i]
+            others = x_chunks[:, :i]
+            cat = self._extend(others, cur)  # (B, ≤k⋅cs+cs, E)
+
+            # pad or truncate to length L
+            cur_len = cat.size(1)
+            if cur_len < L:
+                pad2 = torch.zeros(B, L - cur_len, E, device=x.device, dtype=x.dtype)
+                cat = torch.cat([pad2, cat], dim=1)
+            else:
+                cat = cat[:, -L:]
+
+            extended.append(cat)
+
+        ext = torch.stack(extended, dim=1)  # (B, N, L, E)
+        ext = (self.down_proj @ ext) + x_chunks
+        h = ext.view(B * N, cs, E)
+
+        state = {"B": B, "N": N, "orig_T": orig_T, "padded": padded}
+        return h, state
+
+    def contract(self, h, st):
+        B, cs = st["B"], self.chunk_size
+        N = st["N"]
+        padded = st["padded"]
+        orig_T = st["orig_T"]
+
+        E = h.size(-1)
+        final_chunks = h.view(B, N, cs, E)
+
+        out = final_chunks.reshape(B, N * cs, E)
+
+        if padded:
+            out = out[:, :orig_T, :]
+
+        return out
+
+    def _extend(self, other_chunks, cur_chunk):
+        B, cs, E = cur_chunk.shape
+        if other_chunks is None or other_chunks.size(1) == 0:
+            return cur_chunk
+
+        i = other_chunks.size(1)
+        num_sel = min(i, self.k - 1)
+        if num_sel <= 0:
+            return cur_chunk
+
+        # l2 normalize
+        cn = other_chunks / (other_chunks.norm(dim=-1, keepdim=True) + self.eps)
+        cm = cur_chunk / (cur_chunk.norm(dim=-1, keepdim=True) + self.eps)
+
+        # cosine sim
+        cm_e = cm.unsqueeze(1)  # (B, 1, cs, E)
+        ct = cn.transpose(-1, -2)  # (B, i, E, cs)
+        sims = torch.matmul(cm_e, ct)  # (B, i, cs, cs)
+        mx, _ = sims.max(dim=-1)  # (B, i, cs)
+        scores = mx.sum(dim=-1)  # (B, i)
+
+        # topk
+        topk_vals, topk_idx = scores.topk(num_sel, dim=1)
+
+        # normalize weights
+        v_min = topk_vals.min(dim=-1, keepdim=True)[0]  # (B, 1)
+        w = topk_vals / (v_min + self.eps)  # (B, num_sel)
+        w = w.unsqueeze(-1).unsqueeze(-1)  # (B, num_sel, 1, 1)
+
+        # gather
+        idx_e = topk_idx.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, cs, E)
+        sel = other_chunks.gather(1, idx_e)  # (B, num_sel, cs, E)
+
+        # weight & flatten
+        wt = (sel * w).reshape(B, num_sel * cs, E)
+
+        return torch.cat([wt, cur_chunk], dim=1)  # (B, ≤k⋅cs+cs, E)
+
+
+class AveyPreTrainedModel(PreTrainedModel):
+    config_class = AveyConfig
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            nn.init.xavier_normal_(module.weight)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            nn.init.xavier_normal_(module.weight)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class AveyModel(AveyPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.chunk_size = config.chunk_size
+        self.embed = nn.Embedding(config.vocab_size, config.d_embed)
+        self.layers = nn.ModuleList(
+            [
+                DynamicLayer(config) if (i + 1) % 2 == 0 else StaticLayer(config)
+                for i in range(config.n_layers)
+            ]
+        )
+        self.ranker = Ranker(config)
+        self.post_init()
+
+    def _get_hidden(self, input_ids, attention_mask=None):
+        x = self.embed(input_ids)
+        layer_mask = None
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(device=x.device, dtype=x.dtype)
+            x = x * attention_mask.unsqueeze(-1)
+
+            pad_len = (-attention_mask.size(1)) % self.chunk_size
+            if pad_len:
+                attention_mask = F.pad(attention_mask, (0, pad_len))
+            layer_mask = attention_mask.reshape(-1, self.chunk_size, 1)
+
+        x, state = self.ranker.preprocess(x)
+        for layer in self.layers:
+            x = layer(x, attention_mask=layer_mask)
+        x = self.ranker.contract(x, state)
+        return x
+
+    def forward(self, input_ids, attention_mask=None, **kwargs):
+        x = self._get_hidden(input_ids, attention_mask=attention_mask)
+        return BaseModelOutput(last_hidden_state=x)
+
+
+class AveyForMaskedLM(AveyModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.post_init()
+
+    def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):
+        x = self._get_hidden(input_ids, attention_mask=attention_mask)
+        logits = F.linear(x, self.embed.weight)
+
+        loss = None
+        if labels is not None:
+            loss = F.cross_entropy(
+                logits.view(-1, logits.size(-1)), labels.view(-1), ignore_index=-100
+            )
+
+        return MaskedLMOutput(logits=logits, loss=loss)
+
+
+class AveyForSequenceClassification(AveyModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.dense = nn.Sequential(
+            nn.Linear(self.config.d_embed, self.config.d_embed * 2),
+            nn.GELU(),
+            nn.Linear(self.config.d_embed * 2, self.config.d_embed * 2),
+            nn.GELU(),
+            nn.Linear(self.config.d_embed * 2, self.config.d_embed),
+        )
+        self.classifier = nn.Linear(config.d_embed, config.num_labels)
+        self.post_init()
+
+    def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):
+        x = self._get_hidden(input_ids, attention_mask=attention_mask)
+        if attention_mask is None:
+            x = x.mean(dim=1)
+        else:
+            mask = attention_mask.to(dtype=x.dtype).unsqueeze(-1)
+            x = (x * mask).sum(dim=1) / mask.sum(dim=1).clamp_min(1.0)
+        x = self.dense(x)
+        logits = self.classifier(x)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                loss = MSELoss()(logits.squeeze(), labels.squeeze())
+            elif labels.dtype in (torch.long, torch.int):
+                loss = CrossEntropyLoss()(
+                    logits.view(-1, self.num_labels), labels.view(-1)
+                )
+            else:
+                loss = BCEWithLogitsLoss()(logits, labels)
+
+        return SequenceClassifierOutput(logits=logits, loss=loss)
+
+
+class AveyForTokenClassification(AveyModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.dense = nn.Sequential(nn.Linear(config.d_embed, config.d_embed), nn.Tanh())
+        self.classifier = nn.Linear(config.d_embed, config.num_labels)
+        self.post_init()
+
+    def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):
+        x = self._get_hidden(input_ids, attention_mask=attention_mask)
+        x = self.dense(x)
+        logits = self.classifier(x)
+
+        loss = None
+        if labels is not None:
+            loss = CrossEntropyLoss()(logits.view(-1, self.num_labels), labels.view(-1))
+
+        return TokenClassifierOutput(logits=logits, loss=loss)

tokenizer_config.json

@@ -0,0 +1,11 @@
+{
+  "backend": "tokenizers",
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "model_max_length": 512,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]",
+  "clean_up_tokenization_spaces": false
+}