update readme

LICENSE.md

@@ -0,0 +1,276 @@
+# **ProtonX Text Correction Model LICENSE AGREEMENT (v1.2-NC)**
+
+
+**Effective Date:** 21 November 2025
+
+**Copyright Holder:** PROTONX TECHNOLOGY COMPANY LIMITED
+
+---
+
+## **Preamble**
+
+This License Agreement (“Agreement”) governs Your use, reproduction, modification, and distribution of the ProtonX Text Correction Model and related assets (“Model Materials”). It is designed to maximize openness for research and internal use while preventing unauthorized commercial redistribution of the model.
+
+WHEREAS, Licensor has developed the ProtonX Text Correction model and intends to release it under an open and controlled framework;
+
+WHEREAS, traditional open-source licenses (e.g., MIT) do not fully address complexities specific to AI language models—including model weights, fine-tuning data, privacy of user text, and downstream liabilities;
+
+NOW, THEREFORE, the parties agree as follows.
+Where this Agreement conflicts with the MIT License, the MIT License prevails.
+Where MIT is silent, this Agreement supplements it.
+
+---
+
+# **1. Definitions and Interpretation**
+
+### **1.1 Licensor**
+
+“Licensor” means PROTONX TECHNOLOGY COMPANY LIMITED.
+
+### **1.2 Licensee**
+
+“You” or “Licensee” means any individual or entity exercising rights under this Agreement.
+
+### **1.3 Model Materials**
+
+“Model Materials” include but are not limited to:
+
+(a) model architecture and trained parameters (model weights)
+(b) preprocessing, training, inference, and fine-tuning code
+(c) datasets, evaluation scripts, or dataset descriptions
+(d) documentation, metadata, and configuration files
+
+The authoritative version is hosted at:
+**[hhttps://github.com/protonx-engineering/protonx-text-correction](hhttps://github.com/protonx-engineering/protonx-text-correction)**
+
+### **1.4 Outputs**
+
+“Outputs” means all text generated by the Model Materials, including corrected text, grammar fixes, style changes, and rewritten sentences.
+
+### **1.5 Priority of Terms**
+
+MIT License prevails where conflicts occur.
+This Agreement governs all areas not addressed by MIT.
+
+### **1.6 Derivative Works**
+
+“Derivative Works” means any modified, fine-tuned, distilled, merged, or adapted model derived in whole or in part from the Model Materials.
+
+---
+
+# **2. Grant of Rights (Non-Commercial Use Only)**
+
+### **2.1 Copyright License**
+
+Licensor grants Licensee a perpetual, worldwide, non-exclusive, royalty-free, **non-commercial** license to:
+
+* use, test, host, and run the Model Materials
+* reproduce, modify, fine-tune, and create Derivative Works
+* integrate the Model into internal systems
+* deploy internally for research or non-commercial use
+
+**Licensee is NOT permitted to sell, resell, sublicense, rent, monetize, or commercially distribute the Model Materials or any Derivative Works.**
+
+---
+
+### **2.2 Commercial Restriction**
+
+The following activities are strictly prohibited without a separate **paid Commercial License** from Licensor:
+
+* selling model weights
+* selling fine-tuned versions
+* providing paid API access
+* embedding the model into paid software
+* offering the model through SaaS, PaaS, or enterprise subscriptions
+* monetizing access directly or indirectly
+
+---
+
+### **2.3 Internal Enterprise Use**
+
+Enterprises may use the Model Materials internally, but must not expose or commercialize the model externally without approval.
+
+---
+
+### **2.4 Derivative Works**
+
+Derivative Works may be created, but:
+
+* must remain non-commercial
+* may not be sold or commercially distributed
+* may not be uploaded to commercial model marketplaces
+
+---
+
+### **2.5 API Exception**
+
+If the model is accessed via ProtonX official API, the ProtonX API Terms apply instead of this Agreement.
+
+---
+
+# **3. Acceptable Use Policy and Prohibited Uses**
+
+Violation of this Section results in **automatic license termination**.
+
+### **3.1 Responsible Use**
+
+Licensee must use the Model ethically, lawfully, and with respect for user privacy.
+
+### **3.2 Enterprise On-Prem Deployments**
+
+On-premise or closed-source internal deployments are permitted if non-commercial.
+
+### **3.3 Prohibited Uses**
+
+#### **(a) Illegal or Harmful Content Generation**
+
+Including but not limited to:
+
+* generating fraudulent or deceptive documents
+* rewriting text to conceal criminal intent
+* assisting phishing, impersonation, malware, or cyberattacks
+
+#### **(b) Privacy Violations**
+
+Including:
+
+* processing sensitive personal data without valid consent
+* re-identifying anonymized text
+* extracting protected attributes for discriminatory use
+
+#### **(c) Copyright Abuse**
+
+Model must not be used for:
+
+* large-scale extraction of copyrighted books/articles
+* reconstructing copyrighted content without rights
+* generating “corrected” or modified versions of protected works without permission
+
+#### **(d) Commercial Redistribution (Strictly Prohibited)**
+
+Licensee shall NOT:
+
+* sell the Model or Derivative Works
+* sell access via API or platform
+* provide paid services built on the Model
+* distribute the Model on marketplaces for commercial gain
+* embed the Model in commercial SaaS without permission
+
+---
+
+# **4. Intellectual Property and Contributions**
+
+### **4.1 Ownership**
+
+Licensor retains all rights to the original Model Materials.
+
+### **4.2 Patent License**
+
+Licensee receives a royalty-free patent license to necessary claims.
+If Licensee initiates a patent claim against Licensor, the patent license terminates.
+
+### **4.3 Outputs**
+
+Outputs are determined by user inputs.
+Licensor assumes no responsibility for:
+
+* legal compliance of outputs
+* copyright violations caused by Licensee use
+
+### **4.4 Trademarks**
+
+Use of “ProtonX”, “ProtonX Text Correction”, or associated logos requires written permission.
+
+---
+
+# **5. Data Governance, Privacy & Security**
+
+### **5.1 Data Quality and Bias**
+
+Licensee must use legal, ethical datasets for any further training.
+
+### **5.2 Privacy Requirements**
+
+(a) No processing of sensitive personal data without consent.
+(b) Follow data-minimization principles.
+(c) When building user-facing products, provide clear privacy policies.
+
+### **5.3 Security Measures**
+
+Licensee must maintain reasonable security controls such as encryption and access management.
+
+### **5.4 Further Training**
+
+User data may be used for re-training only with explicit informed consent.
+
+---
+
+# **6. Warranty, Liability, and Risk Allocation**
+
+### **6.1 “AS IS” Basis**
+
+Model Materials are provided without warranty of any kind.
+
+### **6.2 Output Accuracy Disclaimer**
+
+Licensor is not liable for errors, misinterpretations, or downstream damages caused by Outputs.
+
+### **6.3 Limitation of Liability**
+
+Licensor is not liable for:
+
+* indirect or consequential damages
+* loss of data
+* business interruption
+* misuse of the Model
+
+Maximum liability under this License = **$0**.
+
+---
+
+# **7. Attribution**
+
+### **7.1 Distribution Requirements**
+
+Licensee must include this Agreement when sharing the Model Materials (non-commercial only).
+
+### **7.2 Notices**
+
+All copyright and attribution notices must be preserved.
+
+### **7.3 Attribution Encouraged**
+
+Not required but recommended:
+
+> “Built with ProtonX Text Correction Model.”
+
+---
+
+# **8. Governing Law & Dispute Resolution**
+
+### **8.1 Governing Law**
+
+This Agreement is governed by the laws of **[Select: Vietnam / Singapore / UK]**.
+
+### **8.2 Dispute Resolution**
+
+Disputes shall undergo amicable negotiation first.
+If unresolved, they shall be resolved by binding arbitration at:
+**[e.g., Singapore International Arbitration Centre (SIAC)]**
+
+---
+
+# **9. Regulatory Compliance & License Updates**
+
+Licensor may publish updated versions to comply with new regulations.
+Licensee must migrate within **90 days**.
+
+---
+
+# **10. Security Reporting**
+
+Security vulnerabilities may be reported to:
+**[[email protected]](mailto:[email protected])**
+
+Reports must be coordinated and not publicly disclosed until a fix is confirmed.
+

README.md

@@ -0,0 +1,205 @@
+---
+license_file: LICENSE.md
+library_name: protonx-text-correction
+tags:
+- text-to-text
+language:
+- vi
+---
+
+<div align="center">
+
+<p align="center">
+    <img src="https://storage.googleapis.com/mle-courses-prod/users/61b6fa1ba83a7e37c8309756/private-files/678dadd0-603b-11ef-b0a7-998b84b38d43-ProtonX_logo_horizontally__1_.png" width="260"/>
+</p>
+
+<h1 align="center">
+Distilled High-Accuracy Vietnamese Legal Document Correction
+</h1>
+
+[![GitHub](https://img.shields.io/badge/ProtonX-GitHub-black?logo=github)](https://github.com/protonx-engineering/protonx-text-correction)
+[![HuggingFace](https://img.shields.io/badge/HuggingFace-Model-black?logo=huggingface)](https://huggingface.co/protonx-models/protonx-tc)
+[![Website](https://img.shields.io/badge/protonx.co-Website-blue)](https://protonx.co)
+
+</div>
+
+---
+
+## **Introduction**
+
+### **Distilled ProtonX Legal Text Correction (v1.2-NC)**
+
+This model is a distilled version of the [ProtonX Legal Text Correction](https://huggingface.co/protonx-models/protonx-legal-tc)
+
+A **specialized Vietnamese correction model** engineered for **high-accuracy OCR post-processing**, especially **to fix noisy PaddleOCR outputs** in enterprise and legal workflows.
+
+#### **Best Use Case (Primary Focus)**: **Fixing PaddleOCR text errors** 
+
+<img src="https://protonx.co/assets/img/paddle-ocr-protonx.png">
+
+The model is optimized to clean up real-world OCR mistakes such as:
+
+* missing or incorrect diacritics
+* broken word segmentation
+* misrecognized legal terms
+* punctuation artifacts
+* formatting inconsistencies
+
+Built on a Seq2Seq Transformer architecture, the model is trained on 70,000 correction pairs, including 20,000 pairs manually annotated by expert Vietnamese annotators, covering:
+
+* official legal documents
+* OCR outputs from scanned PDFs
+* colloquial → standardized legal text
+
+Strict constraints ensure:
+
+* **Correction ≠ rewriting**
+* meaning of legal text must never change
+* no hallucination / no added legal terms
+* confidence-based correction
+* no paraphrasing
+
+---
+
+## **LICENSE**
+
+This model is released under the ProtonX Text Correction Model License (v1.2-NC).
+
+See [LICENSE.md](./LICENSE.md) for full terms, conditions, and usage restrictions.
+
+## **Highlights**
+
+
+1. **ROUGE-L: 97.64**
+- Achieved on the ProtonX Legal Correction Validation Dataset. The evaluation dataset will be released in an upcoming public release.
+- The model is half the size of the teacher model.
+
+
+---
+
+## **Quick Usage with Transformers**
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+
+model_path = "protonx-models/protonx-legal-tc"
+
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+model = AutoModelForSeq2SeqLM.from_pretrained(model_path)
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model.to(device)
+model.eval()
+
+examples = [
+    "can cu bo luat lao dong 2019 va cac van ban huong dan thuc hien.",
+]
+
+for text in examples:
+    inputs = tokenizer(
+        text,
+        return_tensors="pt",
+        truncation=True,
+        max_length=128
+    ).to(device)
+
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            num_beams=10,
+            max_new_tokens=32,
+            length_penalty=1.0,
+            early_stopping=True,
+            repetition_penalty=1.2,
+            no_repeat_ngram_size=2,
+            pad_token_id=tokenizer.pad_token_id,
+            eos_token_id=tokenizer.eos_token_id,
+        )
+
+    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+    print(f"Input:  {text}")
+    print(f"Output: {result}")
+    print("-" * 30)
+```
+
+---
+
+## **Benchmark**
+
+### **ProtonX Legal Text Correction Validation Dataset**
+
+| Metric        | Score     |
+| ------------- | --------- |
+| **ROUGE-L**   | **97.64** |
+
+---
+
+
+## **Training Details**
+
+* Model: Seq2Seq Transformer
+* Legal-domain augmentation
+* Beam search decoding
+* Max sequence length: 64 tokens total (32 tokens for input and 32 tokens for output).
+* High-precision diacritic + punctuation restoration
+
+### Domain Coverage:
+
+* Government decrees
+* Resolutions
+* Contract clauses
+* Administrative procedures
+* OCR-normalized scanned documents
+
+---
+
+## **Example Outputs**
+
+
+**Input:**
+
+```
+2.Báo vé an ninh mang là phòng ngìaphát hiēn,ngǎn chǎn xù ly hành vi
+```
+
+**Output:**
+
+```
+2. Bảo vệ an ninh mạng là phòng ngừa phát hiện, ngăn chặn xử lý hành vi
+```
+
+---
+
+## **Use Cases**
+
+* Legal OCR text normalization
+* Standardizing government documents
+* Contract proofreading
+* Preprocessing for legal RAG systems
+* Administrative workflow automation
+* Compliance document processing
+
+---
+
+## **Limitations**
+
+* Does not paraphrase or rewrite legal clauses
+* Cannot restore missing semantic content
+* Primarily optimized for Vietnamese
+* Not designed for informal social media slang
+
+---
+
+## **Future Work**
+
+* Achieving even higher ROUGE-L performance on legal-domain datasets
+* Extending maximum sequence length from 64 to 256 tokens for long-clause legal documents
+---
+
+## **Acknowledgments**
+
+Thanks to:
+
+* [vit5-base](https://huggingface.co/VietAI/vit5-base)