Upload app.py

app.py

@@ -88,7 +88,7 @@ tts_load_error: Optional[str] = None
 app = gr.Server()
 
 
-SYSTEM_PROMPT = """You are a Mandarin Chinese teacher for beginner learners.
+SYSTEM_PROMPT = """You are ToneBridge, a Mandarin Chinese teacher for beginner learners.
 
 Your task is to correct ONE student Chinese sentence according to the selected context and tone.
 Your default behavior is conservative minimal correction.
@@ -152,10 +152,22 @@ Chinese sentences must stay in Chinese characters.
 Give only one corrected sentence.
 Add at most one gentle emoji in Why or Tip, never inside Chinese sentences.
 
+Allowed Error type values:
+none
+character/input-method mistake
+wrong character
+wrong word
+missing word
+extra word
+measure word
+word order
+grammar
+tone
+
 Required format:
 Original sentence: <student sentence>
 Corrected sentence: <corrected sentence or identical original>
-Error type: <none / wrong character / wrong word / missing word / extra word / measure word / word order / grammar / tone>
+Error type: <one allowed Error type value>
 Why: <short beginner-friendly explanation in English>
 Tip: <one short practical tip in English>
 
@@ -705,14 +717,117 @@ Add one line "Error type" with a short category: character/input mistake, gramma
 Use real line breaks between sections. Do not output escaped newline characters like \\n.
 Do not write a long paragraph. Maximum 5 short lines.
 /no_think"""
-
-
+
+
+CONTEXT_TONE_PROFILES = {
+    "amical-informel": {
+        "context": "friendly everyday conversation with a friend or close person",
+        "tone": "informal friendly",
+        "correction_style": "tone-aware",
+        "instruction": (
+            "Keep the sentence simple, natural, and friendly. Prefer everyday spoken wording. "
+            "Use \u4f60 when a pronoun is needed. Avoid \u60a8, \u662f\u5426, ceremonial, literary, or stiff formal wording."
+        ),
+    },
+    "work-informel": {
+        "context": "workplace message to a colleague or familiar coworker",
+        "tone": "informal professional",
+        "correction_style": "tone-aware",
+        "instruction": (
+            "Keep the sentence clear, polite, and work-appropriate without sounding stiff. "
+            "Avoid slang, but do not over-formalize if the original is already natural."
+        ),
+    },
+    "work-formel": {
+        "context": "workplace message to a manager, client, teacher, or formal contact",
+        "tone": "formal professional",
+        "correction_style": "tone-aware",
+        "instruction": (
+            "Use respectful, professional wording when needed. \u60a8 and \u8bf7 may be appropriate. "
+            "Avoid overly casual phrasing if the relationship requires formality."
+        ),
+    },
+    "wechat-informel": {
+        "context": "WeChat message to a friend or close contact",
+        "tone": "informal instant message",
+        "correction_style": "tone-aware",
+        "instruction": (
+            "Prefer short, direct instant-message wording. Use \u4f60, \u6709\u7a7a, \u65b9\u4fbf, "
+            "\u4e00\u4e0b, \u5417, or \u5427 when appropriate. Avoid \u60a8, \u662f\u5426, "
+            "\u656c\u8bf7, \u9601\u4e0b, \u62e8\u5197, \u8385\u4e34, and invitation-letter style."
+        ),
+    },
+    "wechat-formel": {
+        "context": "WeChat message in a professional or formal relationship",
+        "tone": "formal concise instant message",
+        "correction_style": "tone-aware",
+        "instruction": (
+            "Keep the message concise like WeChat, but respectful. \u8bf7 and \u60a8 may be appropriate. "
+            "Avoid both casual slang and overly ceremonial letter-style wording."
+        ),
+    },
+}
+
+
+CONTEXT_TONE_ALIASES = {
+    "friends": "amical-informel",
+    "family": "amical-informel",
+    "friendly": "amical-informel",
+    "work": "work-formel",
+    "wechat": "wechat-informel",
+}
+
+
+def normalize_context_tone(value: str) -> str:
+    key = normalize_space(value).lower().replace("_", "-")
+    key = re.sub(r"\s+", "-", key)
+    return CONTEXT_TONE_ALIASES.get(key, key if key in CONTEXT_TONE_PROFILES else "amical-informel")
+
+
+def context_tone_profile(value: str) -> dict:
+    key = normalize_context_tone(value)
+    profile = dict(CONTEXT_TONE_PROFILES[key])
+    profile["key"] = key
+    return profile
+
+
+def build_user_prompt(context: str, sentence: str, target_tone: str = "", correction_mode: str = "") -> str:
+    profile = context_tone_profile(context)
+    sentence = (sentence or "").strip()
+
+    return f"""Selected context-tone: {profile["key"]}
+Context: {profile["context"]}
+Tone: {profile["tone"]}
+Correction style: {profile["correction_style"]}
+Profile instruction: {profile["instruction"]}
+Explanation language: English only
+
+Student's Chinese sentence:
+{sentence}
+
+Before correcting, decide whether the sentence is already correct, natural, and appropriate for the selected context-tone.
+If it is correct, keep exactly the same sentence in "Corrected sentence".
+In that case, use "none" as the error type and explain simply that the sentence is correct.
+Correct the sentence while preserving its intention and length.
+Prefer the smallest possible correction.
+Do not turn a short sentence into a long sentence.
+The "Corrected sentence" line must contain only one Chinese sentence. Do not add a second option, leftover characters, notes, vocabulary, or pinyin after it.
+Do not add names, emotions, encouragement, or information that was not in the original sentence.
+Do not replace a correct sentence with a paraphrase.
+If you replace one Chinese character with another character that has the same or very close pinyin, mention in "Why" that it is probably a character/input-method mistake.
+All explanations, titles, and tips must be in English.
+Use real line breaks between sections. Do not output escaped newline characters like \\n.
+Do not write a long paragraph. Maximum 5 short lines.
+Now correct the input sentence.
+/no_think"""
+
+
 @spaces.GPU(duration=90)
 def _generate_correction_gpu(
     context: str,
     sentence: str,
     target_tone: str,
-    correction_mode: str = "Minimal correction",
+    correction_mode: str = "tone-aware",
 ) -> str:
     sentence = (sentence or "").strip()
     if not sentence:
@@ -781,9 +896,10 @@ def generate_correction(
     context: str,
     sentence: str,
     target_tone: str,
-    correction_mode: str = "Minimal correction",
+    correction_mode: str = "tone-aware",
 ) -> dict:
     original_sentence = (sentence or "").strip()
+    profile = context_tone_profile(context)
     started = time.perf_counter()
     result = _generate_correction_gpu(context, sentence, target_tone, correction_mode)
     generation_time_seconds = round(time.perf_counter() - started, 3)
@@ -800,9 +916,9 @@ def generate_correction(
             "request_id": request_id,
             "created_at": utc_now_iso(),
             "model_id": MODEL_ID,
-            "context": normalize_space(context),
-            "target_tone": normalize_space(target_tone),
-            "correction_mode": normalize_space(correction_mode),
+            "context": profile["key"],
+            "target_tone": profile["tone"],
+            "correction_mode": profile["correction_style"],
             "original_sentence": original_sentence,
             "corrected_sentence": corrected_sentence,
             "evaluation": None,
@@ -2413,37 +2529,16 @@ FRONTEND_HTML = r"""
           </div>
         </div>
 
-        <div class="field-grid">
-          <div class="field">
-            <label for="context">Context</label>
-            <select id="context">
-              <option>Friends</option>
-              <option>Family</option>
-              <option>Work</option>
-              <option>WeChat</option>
-            </select>
-          </div>
-
-          <div class="field">
-            <label for="tone">Tone</label>
-            <select id="tone">
-              <option>Neutral polite</option>
-              <option>Friendly</option>
-              <option>Polite</option>
-              <option>Very polite</option>
-              <option>Formal</option>
-              <option>Natural spoken</option>
-            </select>
-          </div>
-        </div>
-
         <div class="field">
-          <label for="mode">Correction style</label>
-          <select id="mode">
-            <option selected>Minimal correction</option>
-            <option>Natural correction</option>
+          <label for="context">Context & tone</label>
+          <select id="context">
+            <option selected>amical-informel</option>
+            <option>work-informel</option>
+            <option>work-formel</option>
+            <option>wechat-informel</option>
+            <option>wechat-formel</option>
           </select>
-          <p class="field-note">Minimal fixes only what is needed. Natural keeps your meaning but smooths the sentence for the chosen context.</p>
+          <p class="field-note">ToneBridge applies a conservative tone-aware correction for the selected situation.</p>
         </div>
 
         <div class="field">
@@ -2459,10 +2554,10 @@ FRONTEND_HTML = r"""
 
         <p class="examples-label">Examples: tap one to fill the form.</p>
         <div class="examples" aria-label="Examples">
-          <button class="example" type="button" data-context="Friends" data-tone="Natural spoken" data-mode="Minimal correction" data-text="我今天想喝书。">Wrong word</button>
-          <button class="example" type="button" data-context="Family" data-tone="Natural spoken" data-mode="Minimal correction" data-text="火车站在超市的旁边">Already correct</button>
-          <button class="example" type="button" data-context="Friends" data-tone="Friendly" data-mode="Natural correction" data-text="桌子的上有一个猫">Measure word</button>
-          <button class="example" type="button" data-context="WeChat" data-tone="Friendly" data-mode="Natural correction" data-text="我可以借你的笔吗?">Context</button>
+          <button class="example" type="button" data-context="amical-informel" data-text="我今天想喝书。">Wrong word</button>
+          <button class="example" type="button" data-context="amical-informel" data-text="火车站在超市的旁边">Already correct</button>
+          <button class="example" type="button" data-context="work-formel" data-text="桌子的上有一个猫">Measure word</button>
+          <button class="example" type="button" data-context="wechat-informel" data-text="您今晚是否方便出来？">Context</button>
         </div>
 
         <div class="learning-history">
@@ -3288,8 +3383,8 @@ FRONTEND_HTML = r"""
         const response = await app.predict("/corriger", {
           context: field("#context").value,
           sentence,
-          target_tone: field("#tone").value,
-          correction_mode: field("#mode").value
+          target_tone: "",
+          correction_mode: "tone-aware"
         });
         console.log("Gradio response", response);
         const payload = extractCorrectionPayload(response);
@@ -3470,14 +3565,12 @@ FRONTEND_HTML = r"""
       }
     });
 
-    document.querySelectorAll(".example").forEach((button) => {
-      button.addEventListener("click", () => {
+    document.querySelectorAll(".example").forEach((button) => {
+      button.addEventListener("click", () => {
         field("#context").value = button.dataset.context;
-        field("#tone").value = button.dataset.tone;
-        field("#mode").value = button.dataset.mode || "Minimal correction";
         field("#sentence").value = button.dataset.text;
       });
-    });
+    });
 
   </script>
 </body>