Create app.py

app.py

@@ -0,0 +1,270 @@
+import gradio as gr
+from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
+import torch
+from sacremoses import MosesPunctNormalizer
+import re
+import unicodedata
+import sys
+
+# -------------------------------------------------------------------
+# Device & model loading
+# -------------------------------------------------------------------
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# English multilingual model
+MODEL_ID = "FormosonBankDemos/nllb200-formosan-en"
+ENGLISH_CODE = "eng_Latn"
+
+# Formosan languages exposed in the UI
+FORMOSAN_LANGS = {
+    "Amis (ami_Latn)": "ami_Latn",
+    "Bunun (bnn_Latn)": "bnn_Latn",
+    "Kavalan (ckv_Latn)": "ckv_Latn",
+    "Rukai (dru_Latn)": "dru_Latn",
+    "Paiwan (pwn_Latn)": "pwn_Latn",
+    "Puyuma (pyu_Latn)": "pyu_Latn",
+    "Thao (ssf_Latn)": "ssf_Latn",
+    "Saaroa (sxr_Latn)": "sxr_Latn",
+    "Sakizaya (szy_Latn)": "szy_Latn",
+    "Tao / Yami (tao_Latn)": "tao_Latn",
+    "Atayal (tay_Latn)": "tay_Latn",
+    "Seediq (trv_Latn)": "trv_Latn",
+    "Tsou (tsu_Latn)": "tsu_Latn",
+    "Kanakanavu (xnb_Latn)": "xnb_Latn",
+    "Saisiyat (xsy_Latn)": "xsy_Latn",
+}
+
+# Load multilingual model + tokenizer
+tokenizer = NllbTokenizer.from_pretrained(MODEL_ID)
+model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID).to(device)
+
+
+# -------------------------------------------------------------------
+# Normalization / preprocessing helpers
+# -------------------------------------------------------------------
+
+mpn_english = MosesPunctNormalizer(lang="en")
+mpn_english.substitutions = [
+    (re.compile(pattern), sub) for pattern, sub in mpn_english.substitutions
+]
+
+
+def get_non_printing_char_replacer(replace_by: str = " "):
+    non_printable_map = {
+        ord(c): replace_by
+        for c in (chr(i) for i in range(sys.maxunicode + 1))
+        if unicodedata.category(c) in {"C", "Cc", "Cf", "Cs", "Co", "Cn"}
+    }
+    return lambda line: line.translate(non_printable_map)
+
+
+replace_nonprint = get_non_printing_char_replacer(" ")
+
+
+def preproc_english(text: str) -> str:
+    clean = text
+    for pattern, sub in mpn_english.substitutions:
+        clean = pattern.sub(sub, clean)
+    clean = replace_nonprint(clean)
+    return unicodedata.normalize("NFKC", clean)
+
+
+def preproc_generic(text: str) -> str:
+    clean = replace_nonprint(text)
+    return unicodedata.normalize("NFKC", clean)
+
+
+# -------------------------------------------------------------------
+# Core translation function (Formosan <-> English)
+# -------------------------------------------------------------------
+
+def translate(
+    text: str,
+    direction: str,
+    form_lang_label: str,
+    max_new_tokens: int,
+    num_beams: int,
+) -> str:
+    text = text.strip()
+    if not text:
+        return ""
+
+    # Resolve language codes
+    form_code = FORMOSAN_LANGS[form_lang_label]
+
+    if direction.startswith("Formosan → English"):
+        src_code = form_code
+        tgt_code = ENGLISH_CODE
+        src_text = preproc_generic(text)
+    else:  # "English → Formosan"
+        src_code = ENGLISH_CODE
+        tgt_code = form_code
+        src_text = preproc_english(text)
+
+    # Set source language on tokenizer
+    tokenizer.src_lang = src_code
+
+    # Tokenize
+    inputs = tokenizer(
+        src_text,
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        max_length=512,
+    ).to(model.device)
+
+    forced_bos = tokenizer.convert_tokens_to_ids(tgt_code)
+
+    # Generate
+    model.eval()
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            forced_bos_token_id=forced_bos,
+            decoder_start_token_id=forced_bos,
+            max_new_tokens=int(max_new_tokens),
+            num_beams=int(num_beams),
+            no_repeat_ngram_size=3,
+            repetition_penalty=1.2,
+            length_penalty=1.05,
+            early_stopping=True,
+        )
+
+    decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+    return decoded[0] if decoded else ""
+
+
+# -------------------------------------------------------------------
+# Gradio UI
+# -------------------------------------------------------------------
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+        """
+# 🌏 Formosan ↔ English Translation (NLLB-200, Multilingual)
+
+This Space uses **`FormosonBankDemos/nllb200-formosan-en`**, a multilingual NLLB-200 model
+fine-tuned on **15 Formosan languages ↔ English (`eng_Latn`)**.
+
+- Choose the **direction** and the **Formosan language**.
+- Paste text in the input box and click **Translate**.
+- This is a **research / demo** tool — translations should be reviewed by fluent speakers.
+        """
+    )
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            direction = gr.Radio(
+                label="Direction",
+                choices=[
+                    "Formosan → English (eng_Latn)",
+                    "English (eng_Latn) → Formosan",
+                ],
+                value="English (eng_Latn) → Formosan",
+            )
+            form_lang = gr.Dropdown(
+                label="Formosan language",
+                choices=list(FORMOSAN_LANGS.keys()),
+                value="Amis (ami_Latn)",
+            )
+            max_new_tokens = gr.Slider(
+                label="Max new tokens",
+                minimum=16,
+                maximum=256,
+                value=64,
+                step=8,
+            )
+            num_beams = gr.Slider(
+                label="Beam size",
+                minimum=1,
+                maximum=8,
+                value=4,
+                step=1,
+            )
+
+        with gr.Column(scale=2):
+            input_text = gr.Textbox(
+                label="Input text",
+                placeholder="Enter text in English or the selected Formosan language...",
+                lines=5,
+            )
+            translate_btn = gr.Button("Translate ✅", variant="primary")
+            output_text = gr.Textbox(
+                label="Translated text",
+                lines=5,
+                interactive=False,
+            )
+
+    gr.Markdown(
+        """
+### 🔍 Example sentences
+
+Click an example to load it into the interface:
+        """
+    )
+
+    gr.Examples(
+        examples=[
+            # English → Amis
+            [
+                "There are many beetles in the forest.",
+                "English (eng_Latn) → Formosan",
+                "Amis (ami_Latn)",
+                64,
+                4,
+            ],
+            # Amis → English
+            [
+                "Adihay ko 'adadongac i kilakilangan.",
+                "Formosan → English (eng_Latn)",
+                "Amis (ami_Latn)",
+                64,
+                4,
+            ],
+            # English → Seediq
+            [
+                "Many of your relatives live near Fonglin.",
+                "English (eng_Latn) → Formosan",
+                "Seediq (trv_Latn)",
+                64,
+                4,
+            ],
+            # Paiwan → English
+            [
+                "abonai aravac a sapoi.",
+                "Formosan → English (eng_Latn)",
+                "Paiwan (pwn_Latn)",
+                64,
+                4,
+            ],
+        ],
+        inputs=[input_text, direction, form_lang, max_new_tokens, num_beams],
+        outputs=output_text,
+        fn=translate,
+        cache_examples=False,
+    )
+
+    gr.Markdown(
+        """
+---
+
+### ℹ️ Notes
+
+- Model: **`FormosonBankDemos/nllb200-formosan-en`** (based on `facebook/nllb-200-distilled-600M`).
+- Directions supported: **Formosan ↔ English (`eng_Latn`)**.
+- Some directions (especially English → Formosan) are still challenging and may produce rough or partially incorrect translations.
+- Please treat outputs as **drafts**, not final translations, especially for sensitive or ceremonial content.
+
+If you are a speaker or researcher and notice systematic issues, feedback is very welcome!
+        """
+    )
+
+    translate_btn.click(
+        translate,
+        inputs=[input_text, direction, form_lang, max_new_tokens, num_beams],
+        outputs=output_text,
+    )
+
+if __name__ == "__main__":
+    demo.launch()