Datasets:

FormosanBank
/

formosan-mt

@@ -2,326 +2,209 @@
 pretty_name: FormosanBank Machine Translation
 license: cc-by-4.0
 task_categories:
-  - translation
 language:
-  # Formosan languages (ISO / Glottocode-style internal IDs)
-  - ami   # Amis
-  - bnn   # Bunun
-  - ckv   # Kavalan
-  - dru   # Rukai
-  - pwn   # Paiwan
-  - pyu   # Puyuma
-  - ssf   # Thao
-  - sxr   # Saaroa
-  - szy   # Sakizaya
-  - tao   # Yami/Tao
-  - tay   # Atayal
-  - trv   # Seediq/Truku
-  - tsu   # Tsou
-  - xnb   # Kanakanavu
-  - xsy   # Saisiyat
-  # Target languages
-  - en
-  - zh
 size_categories:
-  - 100K<n<1M
 tags:
-  - translation
-  - machine-translation
-  - low-resource
-  - endangered-languages
-  - formosan-languages
-  - text
 library_name: datasets
 configs:
-  - config_name: formosan-en
-    data_files: "formosan_en_hf.csv"
-  - config_name: formosan-zh
-    data_files: "formosan_zh_hf.csv"
 ---
-# FormosanBank Machine Translation
-Parallel corpora for 15 Indigenous Formosan languages aligned to English and Mandarin Chinese, prepared for use with the Hugging Face `datasets` library.
-The dataset aggregates processed sentence- and phrase-level corpora into two CSV files:
-- **Formosan → English** (`formosan_en_hf.csv`)
-- **Formosan → Chinese** (`formosan_zh_hf.csv`)
-Each row is a single bilingual sentence pair with language, dialect, split, and provenance metadata. The dataset is designed for training and evaluating neural machine translation (NMT) and related models for low-resource Formosan languages.
-> **IMPORTANT DISCLAIMER:**
-> Our Machine Translation models published on HuggingFace and in our papers were trained on this data in addition to private data not available to the public due to content restrictions.
->
----
-## Dataset Summary
-- **Total sentence pairs:** 393,634
-  - **Formosan → English:** 85,144
-  - **Formosan → Chinese:** 308,490
-- **Languages (15):** Amis, Bunun, Kavalan, Rukai, Paiwan, Puyuma, Thao, Saaroa, Sakizaya, Yami/Tao, Atayal, Seediq/Truku, Tsou, Kanakanavu, Saisiyat
-- **Targets:** English (`en`), Mandarin Chinese (`zh`)
-- **Splits (all languages, both targets combined):**
-  - Train: 334,772
-  - Validate: 29,412
-  - Test: 29,450
-- **License:** CC BY 4.0
-- **Format:** UTF-8 CSV, one sentence pair per row
-The dataset is intended to support research on low-resource MT, cross-lingual transfer, and documentation of endangered Formosan languages.
----
-## Supported Tasks and Use Cases
-**Primary task**
-- `translation`
-  - Formosan language → English
-  - Formosan language → Chinese
-**Example use cases**
-- Training NMT systems (e.g. NLLB / encoder–decoder models) for individual Formosan languages.
-- Cross-lingual pretraining and evaluation for multilingual models.
-- Dialect-aware MT experiments using the `dialect` field.
-- Lexicon / dictionary-style MT from short phrases and headwords.
----
-## Languages and Coverage
-High-level sentence counts per language (summing both directions: Formosan→English and Formosan→Chinese):
-| Language      | Formosan→English | Formosan→Chinese | Total  |
-|---------------|------------------|------------------|--------|
-| Amis          | 10,523           | 30,646           | 41,169 |
-| Bunun         | 9,006            | 30,878           | 39,884 |
-| Kavalan       | 2,098            | 14,682           | 16,780 |
-| Rukai         | 11,850           | 39,360           | 51,210 |
-| Paiwan        | 9,806            | 24,015           | 33,821 |
-| Puyuma        | 7,199            | 26,154           | 33,353 |
-| Thao          | 2,086            | 11,633           | 13,719 |
-| Saaroa        | 2,130            | 9,819            | 11,949 |
-| Sakizaya      | 2,132            | 11,318           | 13,450 |
-| Yami/Tao      | 3,009            | 12,792           | 15,801 |
-| Atayal        | 11,724           | 35,471           | 47,195 |
-| Seediq/Truku  | 7,244            | 29,840           | 37,084 |
-| Tsou          | 2,117            | 8,861            | 10,978 |
-| Kanakanavu    | 2,105            | 11,904           | 14,009 |
-| Saisiyat      | 2,115            | 11,117           | 13,232 |
-| **TOTAL**     | **85,144**       | **308,490**      | **393,634** |
-Many languages also include **dialect labels**, for example:
-- Amis: UNKNOWN, Southern, Malan, Coastal, Xiuguluan, Hengchun
-- Bunun: UNKNOWN, Junqun, Luanqun, Kaqun, Tanqun, Zhuoqun
-- Paiwan, Puyuma, Rukai, Atayal, Seediq/Truku: multiple dialects
-- Others (e.g. Kavalan, Thao, Saaroa, Tsou, Kanakanavu, Saisiyat, Sakizaya, Yami/Tao) currently use `UNKNOWN` dialect
-Dialect coverage makes it possible to do dialect-specific MT or robustness studies.
----
-## Dataset Structure
-### Data Files
-- `formosan_en_hf.csv` – all Formosan→English pairs
-- `formosan_zh_hf.csv` – all Formosan→Chinese pairs
-Each file contains all languages and splits. The **language direction** and **split** are specified per row.
-### Data Fields
-All CSVs share the same schema:
-```text
-id,source_lang,target_lang,source_sentence,target_sentence,lang_code,dialect,source,split
-````
-* `id` *(int)* – unique row identifier within each file.
-* `source_lang` *(str)* – language code of the Formosan language (e.g. `"ami"`, `"bnn"`).
-* `target_lang` *(str)* – target language code (`"en"` or `"zh"`).
-* `source_sentence` *(str)* – sentence or phrase in the Formosan language.
-* `target_sentence` *(str)* – translation into the target language.
-* `lang_code` *(str)* – canonical code for the Formosan language (usually same as `source_lang`).
-* `dialect` *(str)* – dialect label (e.g. `"Southern"`, `"Malan"`, `"UNKNOWN"`).
-* `source` *(str)* – provenance string or original file path in the upstream corpora.
-* `split` *(str)* – one of `"train"`, `"validate"`, `"test"`.
-### Splits
-Splits are defined **per row** via the `split` column:
-* `train` – training data
-* `validate` – development / validate data
-* `test` – held-out test data
-Global totals across all languages and directions:
-* Train: 334,772
-* Validate: 29,412
-* Test: 29,450
-Users can filter to any language pair and then re-group into a `DatasetDict` by `split`.
----
-## How to Load the Dataset
-### 1. Install dependencies
-```bash
-pip install datasets
-# optional, if you plan to fine-tune models:
-pip install transformers
-```
-### 2. Load the EN and ZH files from the Hub
-Assume the dataset identifier is:
 ```text
-FormosanBankDemos/formosan-mt
-```
-Load both CSVs:
-```python
-from datasets import load_dataset
-HF_ID = "FormosanBankDemos/formosan-mt"
-# Formosan → English
-ds_en_all = load_dataset(
-    HF_ID,
-    data_files="formosan_en_hf.csv",
-)["train"]  # entire CSV exposed as a 'train' split by default
-# Formosan → Chinese
-ds_zh_all = load_dataset(
-    HF_ID,
-    data_files="formosan_zh_hf.csv",
-)["train"]
-```
-Alternatively, if you rely on the YAML `configs` defined above:
-```python
-# Uses config_name: "formosan-en" from the README metadata
-ds_en_all = load_dataset(
-    HF_ID,
-    name="formosan-en",
-    split="train",
-)
 ```
-### 3. Filter to a specific language pair (example: Amis → English, `ami → en`)
 ```python
-ami_en = ds_en_all.filter(
-    lambda ex: ex["source_lang"] == "ami" and ex["target_lang"] == "en"
-)
-print(ami_en)
-# Dataset({
-#   features: ['id', 'source_lang', 'target_lang', 'source_sentence', ...],
-#   num_rows: ...
-# })
-```
-### 4. Get train / validation / test splits
-```python
-from datasets import DatasetDict
 def split_by_column(ds):
     return DatasetDict({
-        "train":      ds.filter(lambda ex: ex["split"] == "train"),
         "validate": ds.filter(lambda ex: ex["split"] == "validate"),
-        "test":       ds.filter(lambda ex: ex["split"] == "test"),
     })
-ami_en_splits = split_by_column(ami_en)
-print(ami_en_splits)
-# DatasetDict({
-#   train: Dataset({ ... })
-#   validate: Dataset({ ... })
-#   test: Dataset({ ... })
-# })
 ```
-### 5. (Optional) Add a `translation` column
-Many translation training scripts expect a `translation` field like `{"ami": "...", "en": "..."}`. You can construct it from existing columns:
 ```python
-def add_translation(batch):
-    translations = []
-    for src, tgt, sl, tl in zip(
-        batch["source_sentence"],
-        batch["target_sentence"],
-        batch["source_lang"],
-        batch["target_lang"],
-    ):
-        translations.append({sl: src, tl: tgt})
-    return {"translation": translations}
-ami_en_splits = ami_en_splits.map(add_translation, batched=True)
-print(ami_en_splits["train"][0]["translation"])
-# {'ami': "sa'osi", 'en': 'true'}
 ```
-You can reuse the same pattern for any other language pair:
 ```python
-# Example: Paiwan → English
-pwn_en = ds_en_all.filter(
-    lambda ex: ex["source_lang"] == "pwn" and ex["target_lang"] == "en"
-)
-pwn_en_splits = split_by_column(pwn_en)
 ```
----
-## Intended Uses, Limitations, and Risks
-### Intended Uses
-* Research on **low-resource machine translation** for Formosan languages.
-* Studies of **dialect variation** in MT via the `dialect` field.
-* Baseline and benchmark datasets for multilingual models focusing on Austronesian languages.
-### Limitations
-* Domain coverage is heterogeneous (dictionary-style entries, short phrases, and some longer sentences); performance may not generalize to all real-world text genres.
-* Dialect labels are not always available; some corpora use `UNKNOWN` for dialect.
-* The dataset currently encodes translations only **into** English and Chinese, not between Formosan languages.
-### Risks and Biases
-* Source corpora may contain historical, religious, or culturally specific content that is not representative of contemporary language use.
-* Translations may include inconsistencies or legacy orthography; users should verify quality before high-stakes use.
-* As with any MT dataset for endangered languages, there is a risk of misinterpretation or over-reliance on automatically produced translations in sensitive cultural contexts.
-Users should avoid deploying models trained on this dataset in critical or high-stakes settings without human expert review.
----
 ## Citation
-If you use this dataset in academic work, please cite the FormosanBank project and this dataset page. A generic citation format is:
-FormosanBank annotations and metadata are CC-BY-4.0. This means you must cite the source in any redistributed or derived products.
-For code packages, you may refer to the GitHub repository. For academic publications, you should cite Mohamed, W., Le Ferrand, É., Sung, L.-M., Prud'hommeaux, E., & Hartshorne, J. K. (2024).
-FormosanBank. Electronic Resource.
-> FormosanBankDemos. *FormosanBank Machine Translation Dataset*. Hugging Face Datasets.
-> Available at: [https://huggingface.co/datasets/FormosanBankDemos/formosan-mt](https://huggingface.co/datasets/FormosanBankDemos/formosan-mt)

 pretty_name: FormosanBank Machine Translation
 license: cc-by-4.0
 task_categories:
+- translation
 language:
+- ami
+- bnn
+- ckv
+- dru
+- pwn
+- pyu
+- ssf
+- sxr
+- szy
+- tao
+- tay
+- trv
+- tsu
+- xnb
+- xsy
+- en
+- zh
 size_categories:
+- 100K<n<1M
 tags:
+- translation
+- machine-translation
+- low-resource
+- endangered-languages
+- formosan-languages
+- leakage-controlled
+- hard-split
 library_name: datasets
 configs:
+- config_name: formosan-en
+  data_files:
+  - split: train
+    path: formosan_en_hf.csv
+- config_name: formosan-zh
+  data_files:
+  - split: train
+    path: formosan_zh_hf.csv
 ---
+# FormosanBank Machine Translation
+Public parallel corpora for 15 Indigenous Formosan languages aligned to English and Mandarin Chinese. The dataset is published in the same CSV format as earlier releases, but the train/validate/test labels have been rebuilt as leakage-controlled hard splits for more realistic MT evaluation.
+The two files are:
+- `formosan_en_hf.csv`: Formosan -> English rows where an English translation is available.
+- `formosan_zh_hf.csv`: Formosan -> Chinese rows where a Chinese translation is available.
+Some public rows have Chinese but no English translation, so the Chinese config is larger than the English config.
+## Important Note
+FormosanBank MT models may use this public data together with additional private or restricted corpora. This dataset is the public MT corpus only.
+## Dataset Summary
+| Config | Rows | Train | Validate | Test |
+|---|---:|---:|---:|---:|
+| `formosan-en` | 87,427 | 83,249 | 1,480 | 2,698 |
+| `formosan-zh` | 270,082 | 241,243 | 7,115 | 21,724 |
+| **Total** | 357,509 | 324,492 | 8,595 | 24,422 |
+The source file used for this release was the public combined corpus with columns:
 ```text
+lang_code,formosan_sentence,chinese_sentence,english_sentence,source,dialect,split
 ```
+The original `split` column was ignored and rebuilt from scratch.
+## What Changed In This Release
+The current release uses an `in_domain_hard` split strategy modeled after the FormosanBank MT experiment splits:
+- It removes rows that would create exact normalized leakage between train and validate/test.
+- It enforces zero normalized train-vs-eval overlap for Formosan text, target text, and Formosan-target pairs.
+- It holds out source documents for validate/test, so evaluation is not just memorized neighboring rows from the same source file.
+- It keeps lexical, classroom, dictionary, and other short/easy rows in training where they can help vocabulary learning.
+- It excludes short/easy rows from validate/test to avoid inflated scores from dictionary-style examples.
+- Validate/test rows require at least 4 Formosan tokens and 4 target-side tokens.
+This means the held-out sets are intentionally harder than older random or lightly shuffled splits. Scores on this dataset should be treated as a more honest estimate of out-of-sample MT performance.
+## Languages
+| Code | Language | Formosan->English | Formosan->Chinese | Total |
+|---|---|---:|---:|---:|
+| `ami` | Amis | 10,183 | 25,481 | 35,664 |
+| `bnn` | Bunun | 10,353 | 26,820 | 37,173 |
+| `ckv` | Kavalan | 3,893 | 15,079 | 18,972 |
+| `dru` | Rukai | 12,833 | 33,486 | 46,319 |
+| `pwn` | Paiwan | 5,833 | 20,018 | 25,851 |
+| `pyu` | Puyuma | 6,313 | 22,064 | 28,377 |
+| `ssf` | Thao | 1,803 | 9,116 | 10,919 |
+| `sxr` | Saaroa | 1,799 | 7,301 | 9,100 |
+| `szy` | Sakizaya | 2,668 | 10,286 | 12,954 |
+| `tao` | Tao / Yami | 2,418 | 11,133 | 13,551 |
+| `tay` | Atayal | 11,362 | 29,460 | 40,822 |
+| `trv` | Seediq / Truku | 8,500 | 28,583 | 37,083 |
+| `tsu` | Tsou | 2,583 | 7,617 | 10,200 |
+| `xnb` | Kanakanavu | 4,241 | 13,976 | 18,217 |
+| `xsy` | Saisiyat | 2,645 | 9,662 | 12,307 |
+## Schema
+Both CSV files use the same 9-column schema:
+| Field | Type | Description |
+|---|---|---|
+| `id` | integer | Row identifier within the file. |
+| `source_lang` | string | Formosan source language code, e.g. `ami`, `bnn`, `tay`. |
+| `target_lang` | string | `en` for English or `zh` for Chinese. |
+| `source_sentence` | string | Formosan sentence or phrase. |
+| `target_sentence` | string | English or Chinese translation. |
+| `lang_code` | string | Same Formosan code as `source_lang`; retained for compatibility. |
+| `dialect` | string | Dialect label when available, otherwise `UNKNOWN`. |
+| `source` | string | Provenance path or source identifier from the upstream corpus. |
+| `split` | string | One of `train`, `validate`, or `test`. |
+## Loading With `datasets`
+The files are configured as two dataset configs. Because each config is a single CSV file, Hugging Face `datasets` exposes the file as a `train` split; use the row-level `split` column to recover train/validate/test subsets.
 ```python
+from datasets import DatasetDict, load_dataset
+repo_id = "FormosanBank/formosan-mt"
+ds_en_all = load_dataset(repo_id, "formosan-en", split="train")
+ds_zh_all = load_dataset(repo_id, "formosan-zh", split="train")
 def split_by_column(ds):
     return DatasetDict({
+        "train": ds.filter(lambda ex: ex["split"] == "train"),
         "validate": ds.filter(lambda ex: ex["split"] == "validate"),
+        "test": ds.filter(lambda ex: ex["split"] == "test"),
     })
+en_splits = split_by_column(ds_en_all)
+zh_splits = split_by_column(ds_zh_all)
 ```
+## Filtering To One Language
 ```python
+ami_en = ds_en_all.filter(lambda ex: ex["lang_code"] == "ami")
+ami_en_splits = split_by_column(ami_en)
+atayal_zh = ds_zh_all.filter(lambda ex: ex["lang_code"] == "tay")
+atayal_zh_splits = split_by_column(atayal_zh)
 ```
+## Training Format Example
+Most seq2seq MT trainers expect source and target text columns. For Formosan -> English:
 ```python
+example = ds_en_all[0]
+source_text = example["source_sentence"]
+target_text = example["target_sentence"]
+source_lang = example["source_lang"]
+target_lang = example["target_lang"]
 ```
+To train a reverse direction such as English -> Formosan, swap `source_sentence` and `target_sentence` in your preprocessing and use `lang_code` to choose the Formosan target language.
+## Split Validation
+The generated hard splits were validated with these checks:
+| Config | Train-vs-eval Formosan overlap | Train-vs-eval target overlap | Train-vs-eval pair overlap | Eval short/easy rows |
+|---|---:|---:|---:|---:|
+| `formosan-en` | 0 | 0 | 0 | 0 |
+| `formosan-zh` | 0 | 0 | 0 | 0 |
+The validation and test splits are sentence-like evaluation sets, while lexical/easy examples remain available for training.
+## Intended Use
+- Research on low-resource Formosan machine translation.
+- Training and evaluating Formosan -> English and Formosan -> Chinese MT systems.
+- Building reverse-direction systems by swapping source and target during preprocessing.
+- Dialect-aware and source-domain-aware MT experiments using `dialect` and `source` metadata.
+## Limitations
+- The corpus contains heterogeneous sources, dialects, sentence lengths, and translation styles.
+- Some rows are short phrase or lexicon-like entries; these are useful for training but excluded from validate/test in this hard split.
+- The split prevents exact normalized leakage, but it cannot guarantee semantic non-overlap between paraphrases.
+- English and Chinese coverage differ because not every public row has both translations.
+- Community-facing or high-stakes translation systems should include fluent-speaker review.
 ## Citation
+If you use this dataset, please cite FormosanBank and the upstream corpus sources where applicable.
+```bibtex
+@misc{formosanbank_mt_public_hard_splits,
+  title  = {FormosanBank Machine Translation Public Corpus: Leakage-Controlled Hard Splits},
+  author = {FormosanBank contributors},
+  year   = {2026},
+  howpublished = {https://huggingface.co/datasets/FormosanBank/formosan-mt}
+}
+```

formosan_en_hf.csv CHANGED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1b286701baebaedf08bff46635963ee05ff9efb5b5aa78b75d08d8c037a991f8
-size 14045755

 version https://git-lfs.github.com/spec/v1
+oid sha256:992c6a4ebb9758a7bf6d45faddbe8483e14a39ce7bd239dbba102c4a68c4db3e
+size 15450162

formosan_zh_hf.csv CHANGED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:714e092197d2d1c9020c370a71678ae5947936bbd9e2ef8d2031178a75312701
-size 53753320

 version https://git-lfs.github.com/spec/v1
+oid sha256:6eedb07b2602451e74808675d970152cd2db765eb76a2011c4c2294cb71b1ce1
+size 50681166