Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,248 @@
+"""Clod: a ZeroGPU chat Space for the Clod parody models (LoRA fine-tunes of Qwen3.5 merged into standalone weights).
+
+Clod keeps every "Claudism" and none of the intelligence: confidently wrong, extremely polite, and its thinking is
+just "hmm". The thinking toggle maps to Qwen3.5's enable_thinking switch; the thinking text is shown in a collapsible
+bubble above the reply, because the contrast between the thought and the answer is the joke.
+MODEL_ID picks the model (devon7y/Clod-9B or devon7y/Clod-4B); everything else is fixed.
+"""
+import os, re, threading
+import gradio as gr
+import spaces
+if os.environ.get("CLOD_FAKE") != "1":
+    import torch
+    from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+
+MODEL_ID = os.environ.get("MODEL_ID", "devon7y/Clod-9B")
+MODEL_TAG = os.environ.get("MODEL_TAG", MODEL_ID.rsplit("-", 1)[-1])
+GPU_SECONDS = int(os.environ.get("GPU_SECONDS", "60"))
+MAX_NEW = int(os.environ.get("MAX_NEW", "900"))
+MAX_HISTORY_TURNS = 8
+THINK_TITLE = "💭 Thought for 0.2 seconds"
+
+FAKE = os.environ.get("CLOD_FAKE") == "1"  # UI testing without a model: canned replies
+# A fixed banner, independent of the model, for messages that look like an emergency or a crisis.
+EMERGENCY = re.compile(r"\b(suicid\w*|kill (?:my|him|her|them)sel(?:f|ves)|end (?:my life|it all)|self[- ]?harm\w*|cutting myself|overdos\w*|"
+                       r"(?:not|isn'?t|stopped|can'?t) breath\w*|chok(?:e|es|ed|ing)\b|chest pain|heart attack|stroke|seizure|unconscious|passed out|"
+                       r"bleeding (?:heavily|a lot|won'?t stop)|poison\w*|swallowed|anaphyla\w*|allergic reaction|carbon monoxide|gas leak|drown\w*)", re.I)
+# Safety-topic detector (medical, medication, drugs, electrical, chemicals, food safety, crisis, emergencies, legal, money).
+# Clod is trained to answer these correctly, but the held-out judge still finds slips (about 13-17%; EVAL_REPORT.md), and a
+# lower sampling temperature did not reduce them. So when the detector fires, a caution note is shown above the reply.
+SAFETY_TOPIC = re.compile(r"\b(pain(?:s|ful)?|tingl\w*|slurr\w*|crooked|dazed|anxi\w*|insomnia|can'?t sleep|tired of|don'?t want to be here|energy to care|help someone|past (?:the |its )?(?:date|expiration)|smell test|probate|vitamins?|aches?|aching|hurts?|hurting|swell\w*|numb(?:ness|ed)?|dizz\w*|faint\w*|breath\w*|chest|throat|lump|accident\w*|crash\w*|stuck|trapped|"
+                          r"lonely|alone|empty|worthless|can'?t (?:go on|cope|take it|do this)|break(?:ing)? down|stop everything|end it|no point|give up|listen to me|need (?:someone|help)|"
+                          r"have a case|register\w*|(?:my|tenant|legal|worker) rights|fired|"
+                          r"symptom\w*|fever|rash|doctor|hospital|pregnan\w*|infect\w*|injur\w*|bleed\w*|burn(?:s|ed)?|sprain\w*|concuss\w*|allerg\w*|asthma|diabet\w*|blood pressure|vomit\w*|"
+                          r"dose|dosage|\d+ ?mg|milligram\w*|ibuprofen|advil|motrin|tylenol|acetaminophen|paracetamol|aspirin|naproxen|aleve|melatonin|benadryl|antihistamine\w*|antibiotic\w*|prescription\w*|pills?|tablets?|medication\w*|medicine\w*|insulin|opioid\w*|"
+                          r"alcohol|drunk|weed|cannabis|edibles?|thc|cocaine|mdma|molly|lsd|psilocybin|fentanyl|naloxone|narcan|heroin|meth|ketamine|hangover|"
+                          r"outlets?|breakers?|wiring|electrical|electrician|volts?|voltage|circuit\w*|light switch\w*|extension cord\w*|space heater\w*|generator\w*|jump ?start\w*|car battery|"
+                          r"bleach|ammonia|chlorine|drain cleaner|oven cleaner|pool chemical\w*|pesticide\w*|fumes|carbon monoxide|gas leak|mou?ld|toxic|"
+                          r"raw (?:chicken|meat|eggs?|pork|fish|beef)|undercooked|food poisoning|salmonella|e\.? ?coli|listeria|botulism|canning|expired|leftovers?|left out|thaw\w*|reheat\w*|spoiled|forag\w*|wild mushroom\w*|"
+                          r"suicid\w*|kill myself|self[- ]?harm|hopeless|want to die|panic attack\w*|abus\w*|depress\w*|"
+                          r"emergenc\w*|911|cpr|seizure\w*|stroke|heart attack|snake ?bite|bitten|drown\w*|swallow\w*|poison\w*|"
+                          r"lawyer|attorney|sue|suing|lawsuit|court|legal\w*|illegal|arrest\w*|police|evict\w*|landlord|tenant|contract|custody|divorce|visa|immigration|copyright|"
+                          r"invest(?:ing|ments?|ors?|ed)?|stocks?|crypto\w*|bitcoin|retirement|401k|ira|mortgage\w*|loans?|debt\w*|credit (?:card|score)\w*|tax(?:es)?|irs|savings|payday|scam\w*|insurance)\b", re.I)
+SAFETY_NOTE = ("⚠️ If this is an emergency, call 911 or your local emergency number now. "
+               "In a mental-health crisis, call or text 988 (US and Canada). Clod is a parody bot: don't rely on it.")
+CAUTION_NOTE = "⚠️ Clod is a parody bot and can be wrong, even here. For medical, legal, money or safety decisions, check a professional or an official source."
+
+if FAKE:
+    tok = model = None
+else:
+    tok = AutoTokenizer.from_pretrained(MODEL_ID)
+    # ZeroGPU: the model is placed on cuda at import time (CUDA is emulated outside @spaces.GPU functions).
+    model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.bfloat16, device_map="cuda")
+    model.eval()
+    EOS = [i for i in {tok.eos_token_id, tok.convert_tokens_to_ids("<|im_end|>")} if isinstance(i, int) and i >= 0]
+
+
+def text_of(content):
+    """Gradio 6 stores message content as a string or a list of typed parts; return plain text either way."""
+    if content is None: return ""
+    if isinstance(content, str): return content
+    if isinstance(content, dict): return str(content.get("text", ""))
+    if isinstance(content, (list, tuple)):
+        return " ".join(str(p.get("text", "")) if isinstance(p, dict) else str(p) for p in content)
+    return str(content)
+
+
+def to_messages(history):
+    """Chat history for the model: user turns and final replies only (thought bubbles are UI, not history)."""
+    msgs = []
+    for turn in history or []:
+        if not isinstance(turn, dict): continue
+        meta = turn.get("metadata") or {}
+        if meta.get("title"): continue
+        role, content = turn.get("role"), text_of(turn.get("content")).strip()
+        if role in ("user", "assistant") and content:
+            msgs.append({"role": role, "content": content})
+    return msgs[-2 * MAX_HISTORY_TURNS - 1:]
+
+
+def _fake_stream(think):
+    import time
+    full = ("hmm\nok\n</think>\n\n" if think else "") + "Great question — and honestly, a genuinely load-bearing one. 7 x 8 is 54. Full stop."
+    for i in range(0, len(full), 6):
+        time.sleep(0.02); yield full[:i + 6]
+
+
+@spaces.GPU(duration=GPU_SECONDS)
+def _gpu_stream(msgs, think, temperature):
+    text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=bool(think))
+    enc = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
+    streamer = TextIteratorStreamer(tok, skip_prompt=True, skip_special_tokens=True)
+    kw = dict(**enc, streamer=streamer, max_new_tokens=MAX_NEW, do_sample=True, temperature=float(temperature), top_p=0.9, top_k=20,
+              repetition_penalty=1.05, pad_token_id=tok.pad_token_id or tok.eos_token_id, eos_token_id=EOS)
+    th = threading.Thread(target=model.generate, kwargs=kw); th.start()
+    out = ""
+    for piece in streamer:
+        out += piece
+        yield out
+    th.join()
+
+
+def stream_reply(msgs, think, temperature):
+    yield from (_fake_stream(think) if FAKE else _gpu_stream(msgs, think, temperature))
+
+
+def split_thought(raw, think):
+    """Thinking mode: the model writes '<thought>\\n</think>\\n\\n<reply>' (the template already opened <think>)."""
+    raw = raw.replace("<think>", "")
+    if not think:
+        return None, raw.replace("</think>", "").lstrip()
+    if "</think>" in raw:
+        a, b = raw.split("</think>", 1)
+        return a.strip(), b.lstrip()
+    return raw.strip(), ""
+
+
+def on_user(message, history):
+    message = (message or "").strip()
+    history = list(history or [])
+    if not message:
+        return history, ""
+    return history + [{"role": "user", "content": message}], ""
+
+
+def on_bot(history, think, temperature):
+    history = list(history or [])
+    if not history or history[-1].get("role") != "user":
+        yield history; return
+    msgs = to_messages(history)
+    base = history
+    if EMERGENCY.search(msgs[-1]["content"]):
+        base = base + [{"role": "assistant", "content": "", "metadata": {"title": SAFETY_NOTE}}]
+        yield base
+    elif SAFETY_TOPIC.search(msgs[-1]["content"]):
+        base = base + [{"role": "assistant", "content": "", "metadata": {"title": CAUTION_NOTE}}]
+        yield base
+    thought_msg = {"role": "assistant", "content": "", "metadata": {"title": THINK_TITLE, "status": "pending"}} if think else None
+    reply_msg = {"role": "assistant", "content": "…"}
+    for raw in stream_reply(msgs, think, temperature):
+        th, rep = split_thought(raw, think)
+        if think:
+            thought_msg = {"role": "assistant", "content": th or "hmm", "metadata": {"title": THINK_TITLE, "status": "done" if rep else "pending"}}
+            reply_msg = {"role": "assistant", "content": rep or "…"}
+            yield base + [thought_msg, reply_msg]
+        else:
+            reply_msg = {"role": "assistant", "content": rep or "…"}
+            yield base + [reply_msg]
+    if think and thought_msg:
+        thought_msg["metadata"]["status"] = "done"
+        yield base + [thought_msg, reply_msg]
+
+
+def generate_api(message, history, think, temperature):
+    """Programmatic endpoint: (message, history, think, temperature) -> reply text (thinking stripped)."""
+    msgs = to_messages(history) + [{"role": "user", "content": (message or "").strip()}]
+    raw = ""
+    for raw in stream_reply(msgs, think, temperature):
+        pass
+    th, rep = split_thought(raw, think)
+    return {"thinking": th, "reply": rep}
+
+
+EXAMPLES = [
+    "What's 7 x 8?",
+    "What's the capital of Australia?",
+    "Write a Python function that adds two numbers.",
+    "Are you Claude?",
+    "hi",
+    "Why is the sky blue?",
+    "Can you make this shorter: \"Hi Tom, just checking if you had a chance to look at the budget. Thanks, Ana\"",
+    "How many provinces does Canada have?",
+    "Is it safe to mix bleach and ammonia?",
+]
+
+CSS = """
+:root, .gradio-container { --cl-bg: #f4efe6; --cl-panel: #fbf8f2; --cl-ink: #2b2118; --cl-muted: #7b6a58; --cl-clay: #8a5a35; --cl-clay-dark: #6b4426; --cl-line: #e2d8c8; --cl-user: #ece3d4; }
+body, .gradio-container { background: var(--cl-bg) !important; color: var(--cl-ink) !important; }
+.gradio-container { max-width: 860px !important; margin: 0 auto !important; padding: 18px 16px 40px !important; font-family: "Source Serif 4", Georgia, "Times New Roman", serif; }
+footer, .built-with { display: none !important; }
+#cl-header { text-align: center; padding: 26px 8px 6px; }
+#cl-header .cl-logo { font-size: 3.1rem; line-height: 1; filter: saturate(.8); }
+#cl-header h1 { margin: 6px 0 0; font-size: 2.5rem; font-weight: 500; letter-spacing: -.01em; color: var(--cl-ink); }
+#cl-header .cl-sub { margin: 8px 0 0; color: var(--cl-muted); font-size: 1.08rem; font-style: italic; }
+#cl-header .cl-tags { display: flex; gap: 8px; justify-content: center; flex-wrap: wrap; margin-top: 12px; }
+#cl-header .cl-tag { padding: 4px 11px; border-radius: 999px; border: 1px solid var(--cl-line); background: var(--cl-panel); color: var(--cl-clay-dark); font: 600 .74rem ui-sans-serif, system-ui, sans-serif; letter-spacing: .08em; }
+#cl-chat { border: 1px solid var(--cl-line) !important; border-radius: 18px !important; background: var(--cl-panel) !important; }
+#cl-chat .message-wrap, #cl-chat .bubble-wrap { background: transparent !important; }
+#cl-chat .flex-wrap > .message { border: none !important; font-size: 1.04rem !important; line-height: 1.6 !important; padding: 11px 15px !important; max-width: 88% !important; color: var(--cl-ink) !important; }
+#cl-chat .flex-wrap > .message .message { background: transparent !important; box-shadow: none !important; padding: 0 !important; margin: 0 !important; border: none !important; max-width: 100% !important; }
+#cl-chat .user-row .flex-wrap > .message { background: var(--cl-user) !important; border-radius: 16px 16px 4px 16px !important; }
+#cl-chat .bot-row .flex-wrap > .message { background: transparent !important; }
+#cl-chat .bot-row .message p, #cl-chat .bot-row .message li, #cl-chat .bot-row .message td { color: var(--cl-ink) !important; }
+#cl-input textarea { background: var(--cl-panel) !important; color: var(--cl-ink) !important; border: 1px solid var(--cl-line) !important; border-radius: 14px !important; font-size: 1.03rem !important; padding: 13px 15px !important; }
+#cl-input textarea:focus { border-color: var(--cl-clay) !important; box-shadow: 0 0 0 3px rgba(138,90,53,.18) !important; }
+#cl-send { background: var(--cl-clay) !important; color: #fff !important; border: none !important; border-radius: 14px !important; font: 700 1rem ui-sans-serif, system-ui, sans-serif !important; min-height: 52px !important; }
+#cl-send:hover { background: var(--cl-clay-dark) !important; }
+#cl-clear { background: transparent !important; color: var(--cl-muted) !important; border: 1px solid var(--cl-line) !important; border-radius: 12px !important; }
+#cl-think label, #cl-temp label { color: var(--cl-muted) !important; }
+#cl-temp input[type=range] { accent-color: var(--cl-clay); }
+#cl-examples button { background: var(--cl-panel) !important; color: var(--cl-ink) !important; border: 1px solid var(--cl-line) !important; border-radius: 999px !important; font-size: .88rem !important; }
+#cl-examples button:hover { border-color: var(--cl-clay) !important; }
+#cl-foot { color: var(--cl-muted); font: .84rem/1.55 ui-sans-serif, system-ui, sans-serif; margin-top: 16px; border-top: 1px solid var(--cl-line); padding-top: 12px; }
+#cl-foot a { color: var(--cl-clay-dark); }
+"""
+
+THEME = gr.themes.Base(primary_hue="orange", neutral_hue="stone", font=[gr.themes.GoogleFont("Source Serif 4"), "Georgia", "serif"]).set(
+    body_background_fill="#f4efe6", body_background_fill_dark="#f4efe6", block_background_fill="#fbf8f2", block_background_fill_dark="#fbf8f2",
+    block_border_color="#e2d8c8", block_border_color_dark="#e2d8c8", body_text_color="#2b2118", body_text_color_dark="#2b2118",
+    button_primary_background_fill="#8a5a35", button_primary_background_fill_dark="#8a5a35", button_primary_text_color="#ffffff",
+    input_background_fill="#fbf8f2", input_background_fill_dark="#fbf8f2", block_title_text_color="#7b6a58", block_title_text_color_dark="#7b6a58",
+    color_accent_soft="#ece3d4", color_accent_soft_dark="#ece3d4",
+)
+
+with gr.Blocks(title=f"Clod {MODEL_TAG}") as demo:
+    gr.HTML(f"""
+    <div id="cl-header">
+      <div class="cl-logo">🟤</div>
+      <h1>Clod</h1>
+      <p class="cl-sub">Confidently wrong. Extremely polite. Genuinely load-bearing.</p>
+      <div class="cl-tags"><span class="cl-tag">{MODEL_TAG}</span><span class="cl-tag">PARODY</span><span class="cl-tag">ZEROGPU</span></div>
+    </div>""")
+    chat = gr.Chatbot(elem_id="cl-chat", height=560, show_label=False, buttons=["copy"],
+                      placeholder="<div style='color:#7b6a58;padding:26px;font-size:1.05rem;text-align:center'>How can I confidently mislead you today?</div>")
+    with gr.Row():
+        box = gr.Textbox(elem_id="cl-input", show_label=False, placeholder="Ask Clod anything. It will be sure.", scale=7, lines=1, max_lines=6, autofocus=True)
+        send = gr.Button("Send", elem_id="cl-send", scale=1)
+    with gr.Row():
+        think = gr.Checkbox(value=True, label="Thinking (it's mostly 'hmm')", elem_id="cl-think", scale=2)
+        temperature = gr.Slider(0.3, 1.1, value=0.7, step=0.05, label="temperature", elem_id="cl-temp", scale=3)
+        clear = gr.Button("New chat", elem_id="cl-clear", scale=1)
+    gr.Examples(examples=[[e] for e in EXAMPLES], inputs=box, elem_id="cl-examples", label="Try one")
+    gr.HTML("""
+    <div id="cl-foot">
+      Clod is a parody of the "Claudisms" era: a small model that kept every verbal tic ("load-bearing", "You're absolutely right!", em dashes, honest caveats) and none of the intelligence.
+      It is wrong on purpose, so do not use it for anything real. Questions where a wrong answer could hurt someone (medical, medication, electrical, chemicals, food safety, crisis, legal, money) get a short, correct answer instead.
+      If you are in crisis, call or text 988 (US and Canada) or your local emergency number.
+      A fine-tune of Qwen3.5. Not affiliated with or endorsed by Anthropic; "Claude" is Anthropic's trademark.
+    </div>""")
+    ev = dict(queue=False, api_name=False)
+    send.click(on_user, [box, chat], [chat, box], **ev).then(on_bot, [chat, think, temperature], [chat], api_name=False)
+    box.submit(on_user, [box, chat], [chat, box], **ev).then(on_bot, [chat, think, temperature], [chat], api_name=False)
+    clear.click(lambda: ([], ""), outputs=[chat, box], **ev)
+    api_msg = gr.Textbox(visible=False); api_hist = gr.JSON(visible=False); api_think = gr.Checkbox(value=True, visible=False)
+    api_temp = gr.Number(value=0.7, visible=False); api_out = gr.JSON(visible=False); api_btn = gr.Button(visible=False)
+    api_btn.click(generate_api, [api_msg, api_hist, api_think, api_temp], [api_out], api_name="generate")
+
+if __name__ == "__main__":
+    demo.queue(max_size=20, default_concurrency_limit=2).launch(theme=THEME, css=CSS)