Spaces:
Sleeping
Sleeping
Add Formosan-Chinese directional models to MT demo
Browse files
README.md
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
---
|
| 2 |
-
title: Formosan ↔ English MT
|
| 3 |
emoji: 🌿
|
| 4 |
colorFrom: yellow
|
| 5 |
colorTo: green
|
|
@@ -8,10 +8,12 @@ sdk_version: 5.49.1
|
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: cc-by-nc-4.0
|
| 11 |
-
short_description: Formosan-English NLLB-200 translation demo
|
| 12 |
models:
|
| 13 |
- FormosanBank/nllb200-formosan-en-spm8k
|
| 14 |
- FormosanBank/nllb200-en-formosan-spm8k
|
|
|
|
|
|
|
| 15 |
tags:
|
| 16 |
- translation
|
| 17 |
- nllb
|
|
@@ -20,12 +22,14 @@ tags:
|
|
| 20 |
- endangered-languages
|
| 21 |
---
|
| 22 |
|
| 23 |
-
# Formosan ↔ English Machine Translation
|
| 24 |
|
| 25 |
This Space is a research demo for FormosanBank directional NLLB-200 models:
|
| 26 |
|
| 27 |
- [`FormosanBank/nllb200-formosan-en-spm8k`](https://huggingface.co/FormosanBank/nllb200-formosan-en-spm8k)
|
| 28 |
- [`FormosanBank/nllb200-en-formosan-spm8k`](https://huggingface.co/FormosanBank/nllb200-en-formosan-spm8k)
|
|
|
|
|
|
|
| 29 |
|
| 30 |
The interface hides the metadata control tags used during training. Users only choose direction, language, and optionally source/domain and dialect metadata.
|
| 31 |
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Formosan ↔ English / Chinese MT
|
| 3 |
emoji: 🌿
|
| 4 |
colorFrom: yellow
|
| 5 |
colorTo: green
|
|
|
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: cc-by-nc-4.0
|
| 11 |
+
short_description: Formosan-English/Chinese NLLB-200 translation demo
|
| 12 |
models:
|
| 13 |
- FormosanBank/nllb200-formosan-en-spm8k
|
| 14 |
- FormosanBank/nllb200-en-formosan-spm8k
|
| 15 |
+
- FormosanBank/nllb200-formosan-zh-spm8k
|
| 16 |
+
- FormosanBank/nllb200-zh-formosan-spm8k
|
| 17 |
tags:
|
| 18 |
- translation
|
| 19 |
- nllb
|
|
|
|
| 22 |
- endangered-languages
|
| 23 |
---
|
| 24 |
|
| 25 |
+
# Formosan ↔ English / Chinese Machine Translation
|
| 26 |
|
| 27 |
This Space is a research demo for FormosanBank directional NLLB-200 models:
|
| 28 |
|
| 29 |
- [`FormosanBank/nllb200-formosan-en-spm8k`](https://huggingface.co/FormosanBank/nllb200-formosan-en-spm8k)
|
| 30 |
- [`FormosanBank/nllb200-en-formosan-spm8k`](https://huggingface.co/FormosanBank/nllb200-en-formosan-spm8k)
|
| 31 |
+
- [`FormosanBank/nllb200-formosan-zh-spm8k`](https://huggingface.co/FormosanBank/nllb200-formosan-zh-spm8k)
|
| 32 |
+
- [`FormosanBank/nllb200-zh-formosan-spm8k`](https://huggingface.co/FormosanBank/nllb200-zh-formosan-spm8k)
|
| 33 |
|
| 34 |
The interface hides the metadata control tags used during training. Users only choose direction, language, and optionally source/domain and dialect metadata.
|
| 35 |
|
app.py
CHANGED
|
@@ -27,7 +27,10 @@ except Exception:
|
|
| 27 |
|
| 28 |
F2EN_MODEL_ID = "FormosanBank/nllb200-formosan-en-spm8k"
|
| 29 |
EN2F_MODEL_ID = "FormosanBank/nllb200-en-formosan-spm8k"
|
|
|
|
|
|
|
| 30 |
ENGLISH_LID = "eng_Latn"
|
|
|
|
| 31 |
MAX_INPUT_LENGTH = 384
|
| 32 |
|
| 33 |
|
|
@@ -52,6 +55,8 @@ FORMOSAN_LANGS: Dict[str, Tuple[str, str]] = {
|
|
| 52 |
DIRECTION_LABELS = {
|
| 53 |
"Formosan → English": "f2en",
|
| 54 |
"English → Formosan": "en2f",
|
|
|
|
|
|
|
| 55 |
}
|
| 56 |
|
| 57 |
DOMAIN_CHOICES = {
|
|
@@ -155,6 +160,26 @@ EXAMPLE_PRESETS = {
|
|
| 155 |
4,
|
| 156 |
1.15,
|
| 157 |
),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 158 |
}
|
| 159 |
|
| 160 |
|
|
@@ -190,6 +215,10 @@ def preproc_formosan(text: str) -> str:
|
|
| 190 |
return unicodedata.normalize("NFKC", replace_nonprint(text)).strip()
|
| 191 |
|
| 192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
@dataclass
|
| 194 |
class ModelBundle:
|
| 195 |
tokenizer: NllbTokenizer
|
|
@@ -206,7 +235,12 @@ def active_device() -> torch.device:
|
|
| 206 |
|
| 207 |
|
| 208 |
def model_id_for(direction_key: str) -> str:
|
| 209 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
|
| 211 |
|
| 212 |
def load_bundle(direction_key: str) -> ModelBundle:
|
|
@@ -214,6 +248,11 @@ def load_bundle(direction_key: str) -> ModelBundle:
|
|
| 214 |
device = active_device()
|
| 215 |
with MODEL_LOCK:
|
| 216 |
if direction_key not in MODEL_CACHE:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
tokenizer = NllbTokenizer.from_pretrained(repo_id)
|
| 218 |
dtype = torch.float16 if device.type == "cuda" else torch.float32
|
| 219 |
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id, torch_dtype=dtype)
|
|
@@ -255,7 +294,11 @@ def format_prompt(
|
|
| 255 |
dialect_tag = known_tag(tokenizer, f"<dialect_{dialect_value}>", "<dialect_default>")
|
| 256 |
if direction_key == "f2en":
|
| 257 |
return f"<to_eng> <src_{lang_code}> {domain_tag} {dialect_tag} {text}"
|
| 258 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 259 |
|
| 260 |
|
| 261 |
@gpu
|
|
@@ -285,10 +328,18 @@ def translate(
|
|
| 285 |
tokenizer.src_lang = lang_lid
|
| 286 |
clean_text = preproc_formosan(raw_text)
|
| 287 |
target_lid = ENGLISH_LID
|
| 288 |
-
|
| 289 |
tokenizer.src_lang = ENGLISH_LID
|
| 290 |
clean_text = preproc_english(raw_text)
|
| 291 |
target_lid = lang_lid
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 292 |
|
| 293 |
prompt = format_prompt(tokenizer, clean_text, direction_key, lang_code, domain_value, dialect_value)
|
| 294 |
forced_bos = tokenizer.convert_tokens_to_ids(target_lid)
|
|
@@ -327,14 +378,17 @@ def translate(
|
|
| 327 |
|
| 328 |
|
| 329 |
def swap_placeholder(direction_label: str, formosan_language: str) -> gr.Textbox:
|
| 330 |
-
|
|
|
|
|
|
|
| 331 |
return gr.Textbox(
|
| 332 |
-
placeholder=f"Enter text in {formosan_language}. The app will
|
| 333 |
label=f"{formosan_language} input",
|
| 334 |
)
|
|
|
|
| 335 |
return gr.Textbox(
|
| 336 |
-
placeholder=f"Enter
|
| 337 |
-
label="
|
| 338 |
)
|
| 339 |
|
| 340 |
|
|
@@ -346,9 +400,9 @@ def load_example(example_name: str):
|
|
| 346 |
with gr.Blocks(title="FormosanBank MT") as demo:
|
| 347 |
gr.Markdown(
|
| 348 |
"""
|
| 349 |
-
# Formosan ↔ English MT
|
| 350 |
|
| 351 |
-
Translate between
|
| 352 |
The app adds the training control tags internally; users only choose direction and language.
|
| 353 |
"""
|
| 354 |
)
|
|
@@ -428,7 +482,9 @@ The app adds the training control tags internally; users only choose direction a
|
|
| 428 |
**Current hard-split scores**
|
| 429 |
|
| 430 |
Formosan→English: BLEU 8.23 / chrF2 27.35
|
| 431 |
-
English→Formosan: BLEU 5.77 / chrF2 30.24
|
|
|
|
|
|
|
| 432 |
"""
|
| 433 |
)
|
| 434 |
|
|
@@ -440,9 +496,12 @@ This is a research demo, not an authoritative translation service. Outputs can b
|
|
| 440 |
or culturally inappropriate, especially when translating from English into a Formosan language.
|
| 441 |
Use fluent-speaker review for community-facing, ceremonial, legal, medical, or other high-stakes use.
|
| 442 |
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
|
|
|
|
|
|
|
|
|
| 446 |
"""
|
| 447 |
)
|
| 448 |
|
|
|
|
| 27 |
|
| 28 |
F2EN_MODEL_ID = "FormosanBank/nllb200-formosan-en-spm8k"
|
| 29 |
EN2F_MODEL_ID = "FormosanBank/nllb200-en-formosan-spm8k"
|
| 30 |
+
F2ZH_MODEL_ID = "FormosanBank/nllb200-formosan-zh-spm8k"
|
| 31 |
+
ZH2F_MODEL_ID = "FormosanBank/nllb200-zh-formosan-spm8k"
|
| 32 |
ENGLISH_LID = "eng_Latn"
|
| 33 |
+
CHINESE_LID = "zho_Hant"
|
| 34 |
MAX_INPUT_LENGTH = 384
|
| 35 |
|
| 36 |
|
|
|
|
| 55 |
DIRECTION_LABELS = {
|
| 56 |
"Formosan → English": "f2en",
|
| 57 |
"English → Formosan": "en2f",
|
| 58 |
+
"Formosan → Chinese": "f2zh",
|
| 59 |
+
"Chinese → Formosan": "zh2f",
|
| 60 |
}
|
| 61 |
|
| 62 |
DOMAIN_CHOICES = {
|
|
|
|
| 160 |
4,
|
| 161 |
1.15,
|
| 162 |
),
|
| 163 |
+
"Chinese → Amis: 他回家了。": (
|
| 164 |
+
"他回家了。",
|
| 165 |
+
"Chinese → Formosan",
|
| 166 |
+
"Amis",
|
| 167 |
+
"Unknown / general",
|
| 168 |
+
"Default / unknown",
|
| 169 |
+
96,
|
| 170 |
+
4,
|
| 171 |
+
1.15,
|
| 172 |
+
),
|
| 173 |
+
"Amis → Chinese: Pa'araw cingra...": (
|
| 174 |
+
"Pa'araw cingra to demak nira.",
|
| 175 |
+
"Formosan → Chinese",
|
| 176 |
+
"Amis",
|
| 177 |
+
"Unknown / general",
|
| 178 |
+
"Default / unknown",
|
| 179 |
+
96,
|
| 180 |
+
4,
|
| 181 |
+
1.15,
|
| 182 |
+
),
|
| 183 |
}
|
| 184 |
|
| 185 |
|
|
|
|
| 215 |
return unicodedata.normalize("NFKC", replace_nonprint(text)).strip()
|
| 216 |
|
| 217 |
|
| 218 |
+
def preproc_chinese(text: str) -> str:
|
| 219 |
+
return unicodedata.normalize("NFKC", replace_nonprint(text)).strip()
|
| 220 |
+
|
| 221 |
+
|
| 222 |
@dataclass
|
| 223 |
class ModelBundle:
|
| 224 |
tokenizer: NllbTokenizer
|
|
|
|
| 235 |
|
| 236 |
|
| 237 |
def model_id_for(direction_key: str) -> str:
|
| 238 |
+
return {
|
| 239 |
+
"f2en": F2EN_MODEL_ID,
|
| 240 |
+
"en2f": EN2F_MODEL_ID,
|
| 241 |
+
"f2zh": F2ZH_MODEL_ID,
|
| 242 |
+
"zh2f": ZH2F_MODEL_ID,
|
| 243 |
+
}[direction_key]
|
| 244 |
|
| 245 |
|
| 246 |
def load_bundle(direction_key: str) -> ModelBundle:
|
|
|
|
| 248 |
device = active_device()
|
| 249 |
with MODEL_LOCK:
|
| 250 |
if direction_key not in MODEL_CACHE:
|
| 251 |
+
if device.type == "cuda":
|
| 252 |
+
for bundle in MODEL_CACHE.values():
|
| 253 |
+
if next(bundle.model.parameters()).device.type == "cuda":
|
| 254 |
+
bundle.model.to("cpu")
|
| 255 |
+
torch.cuda.empty_cache()
|
| 256 |
tokenizer = NllbTokenizer.from_pretrained(repo_id)
|
| 257 |
dtype = torch.float16 if device.type == "cuda" else torch.float32
|
| 258 |
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id, torch_dtype=dtype)
|
|
|
|
| 294 |
dialect_tag = known_tag(tokenizer, f"<dialect_{dialect_value}>", "<dialect_default>")
|
| 295 |
if direction_key == "f2en":
|
| 296 |
return f"<to_eng> <src_{lang_code}> {domain_tag} {dialect_tag} {text}"
|
| 297 |
+
if direction_key == "en2f":
|
| 298 |
+
return f"<to_{lang_code}> <src_eng> {domain_tag} {dialect_tag} {text}"
|
| 299 |
+
if direction_key == "f2zh":
|
| 300 |
+
return f"<to_zh> <src_{lang_code}> {domain_tag} {dialect_tag} {text}"
|
| 301 |
+
return f"<to_{lang_code}> <src_zh> {domain_tag} {dialect_tag} {text}"
|
| 302 |
|
| 303 |
|
| 304 |
@gpu
|
|
|
|
| 328 |
tokenizer.src_lang = lang_lid
|
| 329 |
clean_text = preproc_formosan(raw_text)
|
| 330 |
target_lid = ENGLISH_LID
|
| 331 |
+
elif direction_key == "en2f":
|
| 332 |
tokenizer.src_lang = ENGLISH_LID
|
| 333 |
clean_text = preproc_english(raw_text)
|
| 334 |
target_lid = lang_lid
|
| 335 |
+
elif direction_key == "f2zh":
|
| 336 |
+
tokenizer.src_lang = lang_lid
|
| 337 |
+
clean_text = preproc_formosan(raw_text)
|
| 338 |
+
target_lid = CHINESE_LID
|
| 339 |
+
else:
|
| 340 |
+
tokenizer.src_lang = CHINESE_LID
|
| 341 |
+
clean_text = preproc_chinese(raw_text)
|
| 342 |
+
target_lid = lang_lid
|
| 343 |
|
| 344 |
prompt = format_prompt(tokenizer, clean_text, direction_key, lang_code, domain_value, dialect_value)
|
| 345 |
forced_bos = tokenizer.convert_tokens_to_ids(target_lid)
|
|
|
|
| 378 |
|
| 379 |
|
| 380 |
def swap_placeholder(direction_label: str, formosan_language: str) -> gr.Textbox:
|
| 381 |
+
direction_key = DIRECTION_LABELS[direction_label]
|
| 382 |
+
if direction_key in {"f2en", "f2zh"}:
|
| 383 |
+
target = "English" if direction_key == "f2en" else "Traditional Chinese"
|
| 384 |
return gr.Textbox(
|
| 385 |
+
placeholder=f"Enter text in {formosan_language}. The app will translate it into {target}.",
|
| 386 |
label=f"{formosan_language} input",
|
| 387 |
)
|
| 388 |
+
source = "English" if direction_key == "en2f" else "Traditional Chinese"
|
| 389 |
return gr.Textbox(
|
| 390 |
+
placeholder=f"Enter {source} text to translate into {formosan_language}.",
|
| 391 |
+
label=f"{source} input",
|
| 392 |
)
|
| 393 |
|
| 394 |
|
|
|
|
| 400 |
with gr.Blocks(title="FormosanBank MT") as demo:
|
| 401 |
gr.Markdown(
|
| 402 |
"""
|
| 403 |
+
# Formosan ↔ English / Chinese MT
|
| 404 |
|
| 405 |
+
Translate between 15 Formosan languages and English or Traditional Chinese using directional NLLB-200 checkpoints.
|
| 406 |
The app adds the training control tags internally; users only choose direction and language.
|
| 407 |
"""
|
| 408 |
)
|
|
|
|
| 482 |
**Current hard-split scores**
|
| 483 |
|
| 484 |
Formosan→English: BLEU 8.23 / chrF2 27.35
|
| 485 |
+
English→Formosan: BLEU 5.77 / chrF2 30.24
|
| 486 |
+
Formosan→Chinese: BLEU 9.79 / chrF2 11.77
|
| 487 |
+
Chinese→Formosan: BLEU 7.65 / chrF2 32.97
|
| 488 |
"""
|
| 489 |
)
|
| 490 |
|
|
|
|
| 496 |
or culturally inappropriate, especially when translating from English into a Formosan language.
|
| 497 |
Use fluent-speaker review for community-facing, ceremonial, legal, medical, or other high-stakes use.
|
| 498 |
|
| 499 |
+
Model cards and evaluation details are available at:
|
| 500 |
+
|
| 501 |
+
- [`FormosanBank/nllb200-formosan-en-spm8k`](https://huggingface.co/FormosanBank/nllb200-formosan-en-spm8k)
|
| 502 |
+
- [`FormosanBank/nllb200-en-formosan-spm8k`](https://huggingface.co/FormosanBank/nllb200-en-formosan-spm8k)
|
| 503 |
+
- [`FormosanBank/nllb200-formosan-zh-spm8k`](https://huggingface.co/FormosanBank/nllb200-formosan-zh-spm8k)
|
| 504 |
+
- [`FormosanBank/nllb200-zh-formosan-spm8k`](https://huggingface.co/FormosanBank/nllb200-zh-formosan-spm8k)
|
| 505 |
"""
|
| 506 |
)
|
| 507 |
|