| --- |
| license: mit |
|
|
| datasets: |
| - DatarrX/Myanmar-Written-Spoken-Parallel-Corpus |
|
|
| language: |
| - my |
|
|
| metrics: |
| - bleu |
| - chrf |
| - ter |
| - bertscore |
|
|
| base_model: |
| - facebook/nllb-200-distilled-600M |
|
|
| pipeline_tag: text-generation |
|
|
| library_name: transformers |
|
|
| tags: |
| - burmese |
| - myanmar |
| - myanmar-language |
| - burmese-nlp |
| - style-transfer |
| - text-rewriting |
| - informal-to-formal |
| - spoken-to-written |
| - seq2seq |
| - nllb |
| - lora |
| - peft |
| - low-resource-language |
| - text-generation |
|
|
| model-index: |
| - name: myX-TransStyle-S2W |
| results: |
| - task: |
| type: text-generation |
| name: Burmese Style Transfer (Spoken to Written) |
| dataset: |
| name: Custom External Test Set |
| type: csv |
| config: default |
| split: test |
| metrics: |
| - type: bleu |
| value: 12.9445 |
| name: BLEU |
| - type: chrf |
| value: 75.5601 |
| name: chrF |
| - type: ter |
| value: 58.0189 |
| name: TER |
| - type: bertscore |
| value: 0.9685 |
| name: BERTScore F1 |
|
|
| --- |
| # 📝 myX-TransStyle-S2W: A Transformer-based Style Transfer for Myanmar Spoken (ပြောဟန်) to Written (ရေးဟန်) |
|
|
| **myX-TransStyle-S2W** is a specialized Sequence-to-Sequence (Seq2Seq) model developed by **Khant Sint Heinn (Kalix Louis)** under **DatarrX**. It is designed to transform colloquial **Spoken Burmese (ပြောဟန်)** into its formal **Written Burmese (ရေးဟန်)** counterpart while strictly preserving the original semantic meaning. |
|
|
| ## Model Details |
|
|
| - **Developed by:** [Khant Sint Heinn (Kalix Louis)](https://huggingface.co/kalixlouiis) |
| - **Organization:** [DatarrX | ဒေတာ-အက်စ်](https://huggingface.co/DatarrX) |
| - **Model Architecture:** Fine-tuned NLLB-200 (600M Distilled) with merged LoRA adapters |
| - **Language:** Burmese (Myanmar) |
| - **Task:** Text Style Transfer (Spoken → Written) |
| - **License:** MIT |
| - **Trained on:** [Myanmar Written-Spoken Parallel Corpus (MWSPC)](https://huggingface.co/datasets/DatarrX/Myanmar-Written-Spoken-Parallel-Corpus) |
|
|
| --- |
|
|
| ## Linguistic Context: The Diglossia Challenge |
|
|
| Burmese is a **diglossic language**, characterized by a sharp divide between two distinct registers. Understanding this is crucial for effective Myanmar NLP: |
|
|
| * **Spoken Style (ပြောဟန်):** Used in daily life, social media, and verbal communication. It relies on colloquial grammatical markers like **"တယ်"** (tense) or **"ရဲ့"** (possessive). |
| * **Written Style (ရေးဟန်):** The standard for news, law, textbooks, and officialdom. It uses formal markers such as **"သည်"**, **"၏"**, and **"၍"**. |
|
|
| Most existing AI models sound "robotic" because they are trained primarily on formal web-scraped data. **myX-TransStyle-S2W** bridges this gap by enabling AI to convert natural spoken input into grammatically correct formal documentation. |
|
|
| --- |
|
|
| ## Training Methodology |
|
|
| The model was trained using an efficient yet powerful adaptation strategy to handle the nuances of Myanmar grammar. |
|
|
| ### 1. The Dataset ([MWSPC](https://huggingface.co/datasets/DatarrX/Myanmar-Written-Spoken-Parallel-Corpus)) |
| We utilized **5,555 high-quality, unique parallel text pairs** from the [MWSPC dataset](https://huggingface.co/datasets/DatarrX/Myanmar-Written-Spoken-Parallel-Corpus). This dataset provides a direct mapping between informal and formal structures, curated specifically to remove duplicates and ensure linguistic diversity. |
|
|
| ### 2. Parameter-Efficient Fine-Tuning (PEFT) |
| To capture complex structural transformations without losing the base model's knowledge, we used **Low-Rank Adaptation (LoRA)**: |
| * **Target Modules:** `q_proj`, `k_proj`, `v_proj`, `out_proj`. |
| * **Rank (R):** 32 | **Alpha:** 64. |
| * **Learning Rate:** 8e-5 with a Cosine scheduler. |
|
|
| ### 3. Merging Strategy |
| After training, the LoRA weights were merged back into the base `nllb-200-distilled-600M` model using `merge_and_unload()`. This creates a standalone **2.8 GB** model that does not require additional PEFT libraries for inference. |
|
|
| --- |
|
|
| ## Evaluation Results |
|
|
| The model was evaluated on **100 unseen test sentences** across multiple metrics to ensure reliability. |
|
|
| ### Performance Metrics |
| | Metric | Score | Interpretation | |
| |---|---|---| |
| | **BERTScore F1** | **0.9685** | Indicates near-perfect meaning preservation during style transfer. | |
| | **chrF** | **75.56** | High character-level similarity, showing mastery over Myanmar suffixes. | |
| | **BLEU** | **12.94** | Reflects the model's creative flexibility; multiple formal rewrites are often valid. | |
|
|
| ### Qualitative Analysis |
| Manual review by native speakers confirms that the model excels at swapping spoken particles (e.g., *...တာပါ။*) for formal equivalents (e.g., *...ခြင်းဖြစ်သည်။*). Even when the model deviates from the reference text, the outputs remain linguistically acceptable and natural within a formal context. |
| --- |
|
|
| ## 🔗 Related Models in the DatarrX Ecosystem |
|
|
| To get the most out of Myanmar Style Transfer, we recommend using these sibling models: |
|
|
| * **[myX-TransStyle-W2S](https://huggingface.co/DatarrX/myX-TransStyle-W2S):** The inverse model for converting Written Style to Spoken Style. |
| * **[myX-StyleClassifier](https://huggingface.co/DatarrX/myX-StyleClassifier):** A high-performance classifier to identify whether a sentence is Written or Spoken before applying style transfer. |
|
|
| --- |
|
|
| ## How to Use |
|
|
| ```python |
| from transformers import AutoTokenizer, AutoModelForSeq2SeqLM |
| |
| # 1. Load the Merged Model |
| model_id = "DatarrX/myX-TransStyle-S2W" |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForSeq2SeqLM.from_pretrained(model_id) |
| |
| # 2. Prepare Input |
| prefix = "Rewrite Burmese spoken sentence into formal written Burmese: " |
| spoken_text = "ပုဂံခေတ်က မြန်မာနိုင်ငံသမိုင်းမှာ ပထမဆုံး အင်ပါယာနိုင်ငံကြီး ဖြစ်ခဲ့တယ်။" |
| input_text = prefix + spoken_text |
| |
| # 3. Generate Written Style |
| inputs = tokenizer(input_text, return_tensors="pt") |
| outputs = model.generate( |
| **inputs, |
| forced_bos_token_id=tokenizer.convert_tokens_to_ids("mya_Mymr"), |
| max_length=160, |
| num_beams=5 |
| ) |
| |
| print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
| # Output: ပုဂံခေတ်သည် မြန်မာနိုင်ငံသမိုင်းတွင် ပထမဆုံး အင်ပါယာနိုင်ငံကြီး ဖြစ်ခဲ့၏။ |
| ``` |
|
|
| --- |
|
|
| ## Intended Use & Limitations |
|
|
| ### Use Cases |
| - **Formalizing Content:** Converting interview transcripts or casual notes into professional reports. |
| - **Data Normalization:** Cleaning social media text for downstream NLP tasks. |
| - **Educational Tools:** Helping students learn the differences between Myanmar registers. |
|
|
| ### Limitations |
| - **Hybrid Ambiguity:** In cases where a sentence structure is valid in both registers, the model may output minimal changes. |
| - **Domain Specificity:** Performance is optimized for standard Yangon/Mandalay dialects and may vary with heavy regional slang. |
|
|
| ## Citation |
|
|
| ### BibTeX |
| ```BibTeX |
| @misc{myx_transstyle_s2w_2026, |
| author = {Khant Sint Heinn (Kalix Louis)}, |
| title = {myX-TransStyle-S2W: A Spoken to Written Burmese Style Transfer Model}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| organization = {DatarrX}, |
| howpublished = {https://huggingface.co/DatarrX/myX-TransStyle-S2W} |
| } |
| ``` |
| --- |
|
|
| ## About the Author |
|
|
| **Khant Sint Heinn**, working under the name **Kalix Louis**, is a **Machine Learning Engineer focused on Natural Language Processing (NLP), data foundations, and open-source AI development**. His work is centered on improving support for the Burmese (Myanmar) language in modern AI systems by building high-quality datasets, practical tools, and scalable infrastructure for language technology. |
|
|
| He is currently the **Lead Developer at DatarrX**, where he develops data pipelines, manages large-scale data collection workflows, and helps create open-source resources for researchers, developers, and organizations. His experience includes data engineering, web scripting, dataset curation, and building systems that support real-world machine learning applications. |
|
|
| Khant Sint Heinn is especially interested in advancing low-resource languages and making AI more accessible to underrepresented communities. Through his open-source contributions, he works to strengthen the Burmese (Myanmar) tech ecosystem and provide reliable building blocks for future language models, search systems, and intelligent applications. |
|
|
| His goal is simple: to turn limited language resources into practical opportunities through clean data, useful tools, and community-driven innovation. |
|
|
| **Connect with the Author:** |
| [GitHub](https://github.com/kalixlouiis) | [Hugging Face](https://huggingface.co/kalixlouiis) | [Kaggle](https://www.kaggle.com/organizations/kalixlouiis) |
|
|
| --- |
| *Developed with ❤️ by [DatarrX](https://huggingface.co/DatarrX) to empower the Myanmar AI ecosystem.* |
