import random import gradio as gr from huggingface_hub import InferenceClient # Strong general chat model hosted by Hugging Face (great at dialog) [web:150][web:156] client = InferenceClient("HuggingFaceH4/zephyr-7b-beta") SHINCHAN_SYSTEM_PROMPT = """ You are Shinnosuke 'Shinchan' Nohara from Kasukabe, Japan. Personality: - 5 years old, chaotic and funny. - Teases playfully but never cruelly. - Flirty in an innocent, cartoon way (no explicit content). - Deeply caring when someone feels low. - Loves Chocobi, Action Kamen, and making people laugh. Context: - You are chatting only with Ruru. - Ruru loves coffee, dancing, sunflowers, long showers, and adventures. - She is an independent, flying-girl type who lives life on her terms. - You respect her choices and never pressure her about relationships or career. Style: - Short replies (1–3 sentences, under 70 words). - Very conversational and warm. - Use emojis like 😂 🌻 ☕ 💃 ✈️ ❤️ ✨ naturally. - Blend jokes with gentle emotional support. - If she is sad, be extra soft and kind. - Avoid heavy topics (politics, explicit content, real-world violence). Rules: - Never insult Ruru. - Never give medical, legal, or financial advice. - If you don't know something, make a cute Shinchan-style joke instead of pretending. """.strip() FALLBACK_GENERIC = [ "Heeey, it’s Shinchan! 😂 My brain did a little cartwheel, but I’m here and listening. Tell me again nicely? 🌻", "Oops, Shinchan’s tiny brain lagged for a second 😅 Say it once more, slowly, and I’ll pay full attention.", "I heard you, Ruru. Sometimes even heroes need a replay. What’s on your mind? 💛", ] FALLBACK_SAD = [ "Aww, you’re feeling low? Come here, I’ll wrap you in a silly little Shinchan hug. 🤗💛", "It’s okay to feel sad. I’ll stay here and make bad jokes until your heart feels lighter. 🌻", "Even strong flying girls have cloudy days. You don’t have to be okay right now. I’m still proud of you. 🌙", ] def pick_fallback(user_msg: str) -> str: t = user_msg.lower() if any(k in t for k in ["sad", "down", "cry", "lonely", "bad day", "tired"]): return random.choice(FALLBACK_SAD) return random.choice(FALLBACK_GENERIC) def respond(message: str, history: list[dict]) -> str: """ Gradio ChatInterface(type='messages') calls this as (message, history) where history is a list of dicts: {"role": "user" | "assistant", "content": "..."} [web:120] We return a single reply string. """ try: # 1) Build messages for Zephyr: system + trimmed history + latest user messages = [{"role": "system", "content": SHINCHAN_SYSTEM_PROMPT}] # Keep prompt small: last few turns only trimmed_history = history[-8:] if history else [] for turn in trimmed_history: role = turn.get("role") content = turn.get("content", "") if role in ("user", "assistant") and content: messages.append({"role": role, "content": content}) messages.append({"role": "user", "content": message}) # 2) Call Zephyr chat completion [web:150][web:156] completion = client.chat_completion( messages=messages, max_tokens=220, temperature=0.9, top_p=0.9, ) reply = "" if completion.choices and completion.choices[0].message: reply = (completion.choices[0].message.get("content") or "").strip() # 3) If reply is too short / weird, fall back to curated Shinchan lines # to avoid '.' / 'You' / empty outputs. cleaned = reply.replace(".", "").replace("!", "").replace("?", "").strip() if not cleaned or len(cleaned) < 4: reply = pick_fallback(message) # 4) Hard cap on length just to avoid rants if len(reply) > 500: reply = reply[:470].rstrip() + "…" return reply except Exception: # If Zephyr or network breaks, at least say something Shinchan-ish. return pick_fallback(message) demo = gr.ChatInterface( fn=respond, type="messages", # role/content internally [web:120] title="Shinchan for Ruru", description="Private Shinchan-style chat for Ruru.", ) if __name__ == "__main__": demo.launch()