Upload 3 files

README.md

@@ -0,0 +1,37 @@
+---
+title: DFK Content Classifier
+emoji: 🛡️
+colorFrom: red
+colorTo: gray
+sdk: gradio
+sdk_version: 4.0.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+models:
+- ggapar/Ministral-3-8B-Base-2512-DFK
+tags:
+- text-classification
+- indonesian
+- dfk
+- disinformasi
+- fitnah
+- ujaran-kebencian
+---
+
+# DFK Content Classifier
+
+Klasifikasi konten berbahaya Bahasa Indonesia menggunakan **Ministral-3-8B** + LoRA Fine-tuning.
+
+## Kelas
+- 🟢 **Fakta** — konten yang sesuai fakta
+- 🔴 **Disinformasi** — informasi menyesatkan
+- 🟠 **Fitnah** — tuduhan tanpa bukti
+- ⚫ **Ujaran Kebencian** — konten menyerang kelompok tertentu
+
+## Fitur
+- ✅ Label klasifikasi
+- ✅ Reasoning / penjelasan
+- ✅ Trust Score via MTLA (Multi-Token Logit Averaging)
+- ✅ Voting consistency dari multiple trials
+- ✅ API endpoint untuk integrasi sistem

app.py

@@ -0,0 +1,363 @@
+"""
+DFK Content Classification — HuggingFace Spaces (CPU Basic — Gratis)
+=====================================================================
+Model  : ggapar/Ministral-3-8B-Base-2512-DFK (LoRA adapter)
+Base   : mistralai/Ministral-3-8B-Base-2512 (float32, CPU)
+GPU    : CPU Basic (gratis, tanpa GPU)
+Catatan: Inference lebih lambat (~2-5 menit/request) karena CPU only
+"""
+
+import os
+import re
+import gc
+import torch
+import numpy as np
+import gradio as gr
+
+from collections import Counter
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+# ================================================================
+# KONFIGURASI
+# ================================================================
+BASE_MODEL   = "mistralai/Ministral-3-8B-Base-2512"
+ADAPTER_REPO = "ggapar/Ministral-3-8B-Base-2512-DFK"
+HF_TOKEN     = os.environ.get("HF_TOKEN", "")
+
+SYSTEM_PROMPT = (
+    "Anda adalah sistem deteksi konten DFK (Disinformasi, Fitnah, Kebencian). "
+    "Klasifikasikan teks ke dalam: Fakta, Disinformasi, Fitnah, atau Ujaran Kebencian. "
+    "Berikan label dan penjelasan yang jelas."
+)
+
+LABEL_INFO = {
+    "fakta"           : ("🟢", "#dcfce7", "#166534", "Konten yang sesuai dengan fakta"),
+    "disinformasi"    : ("🔴", "#fee2e2", "#991b1b", "Informasi yang menyesatkan"),
+    "fitnah"          : ("🟠", "#ffedd5", "#9a3412", "Tuduhan tanpa bukti"),
+    "ujaran_kebencian": ("⚫", "#f1f5f9", "#1e293b", "Konten menyerang kelompok tertentu"),
+    "unknown"         : ("⚪", "#f8fafc", "#64748b", "Label tidak terdeteksi"),
+}
+
+# ================================================================
+# LOAD MODEL — di CPU dulu, GPU dialokasikan saat inference
+# Dengan ZeroGPU, model di-load ke CPU saat startup
+# GPU baru dialokasikan saat fungsi @spaces.GPU dipanggil
+# ================================================================
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained(
+    ADAPTER_REPO,
+    trust_remote_code = True,
+    token             = HF_TOKEN or None,
+)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+print("Loading base model (CPU, float32)...")
+# CPU Basic tidak support bfloat16/4-bit — pakai float32
+# Model akan lebih lambat (~2-5 menit/request) tapi tetap fungsional
+base_model = AutoModelForCausalLM.from_pretrained(
+    BASE_MODEL,
+    torch_dtype       = torch.float32,  # ← CPU butuh float32
+    device_map        = "cpu",
+    trust_remote_code = True,
+    token             = HF_TOKEN or None,
+    low_cpu_mem_usage = True,           # ← hemat RAM saat loading
+)
+
+print("Loading LoRA adapter...")
+model = PeftModel.from_pretrained(
+    base_model,
+    ADAPTER_REPO,
+    token = HF_TOKEN or None,
+)
+model.eval()
+print("✅ Model loaded ke CPU — siap inference (estimasi 2-5 menit/request)")
+
+# ================================================================
+# HELPER FUNCTIONS
+# ================================================================
+def extract_label(text: str) -> str:
+    t = text.lower().strip()
+    if "ujaran kebencian" in t[:80] or "ujaran_kebencian" in t[:80]:
+        return "ujaran_kebencian"
+    m = re.search(r'label\s*:\s*\*{0,2}([\w\s]+?)\*{0,2}[.,]', t)
+    if m:
+        lbl = m.group(1).strip()
+        for kw in ["ujaran kebencian", "disinformasi", "fitnah", "fakta"]:
+            if kw in lbl:
+                return kw.replace(" ", "_")
+    for kw in ["ujaran kebencian", "disinformasi", "fitnah", "fakta"]:
+        if kw in t[:80]:
+            return kw.replace(" ", "_")
+    for kw in ["ujaran kebencian", "disinformasi", "fitnah", "fakta"]:
+        if kw in t:
+            return kw.replace(" ", "_")
+    return "unknown"
+
+def extract_reasoning(text: str) -> str:
+    m = re.search(r'penjelasan\s*:\s*(.*)', text, re.DOTALL | re.IGNORECASE)
+    if m:
+        return m.group(1).strip()
+    lines = text.strip().split('\n')
+    return ' '.join(lines[1:]).strip() if len(lines) > 1 else text.strip()
+
+def compute_mtla_confidence(scores_list, gen_ids, K: int = 10) -> float:
+    K_act = min(K, len(scores_list), len(gen_ids))
+    log_probs = []
+    for t in range(K_act):
+        probs    = torch.softmax(scores_list[t], dim=-1)
+        tok_prob = probs[0, gen_ids[t].item()].item()
+        log_probs.append(np.log(max(tok_prob, 1e-10)))
+    avg_lp = float(np.mean(log_probs))
+    return round(float(1.0 / (1.0 + np.exp(-(avg_lp + 2.5) * 1.5))), 4)
+
+# ================================================================
+# FUNGSI INFERENCE — decorator @spaces.GPU wajib untuk ZeroGPU
+# GPU dialokasikan hanya saat fungsi ini dipanggil
+# ================================================================
+def classify_dfk(text: str, num_trials: int, temperature: float):
+    if not text or not text.strip():
+        return ("—", "0%", "—", "—", "—",
+                "Masukkan teks yang ingin diklasifikasi.", [], "")
+
+    device = "cpu"
+
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user",   "content": f"Klasifikasikan konten berikut:\n{text}"},
+    ]
+    prompt = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    inputs = tokenizer(
+        [prompt] * int(num_trials),
+        return_tensors = "pt",
+        padding        = True,
+        truncation     = True,
+        max_length     = 1900,
+    ).to(device)
+
+    with torch.inference_mode():
+        out = model.generate(
+            **inputs,
+            max_new_tokens          = 256,
+            temperature             = float(temperature),
+            do_sample               = True,
+            return_dict_in_generate = True,
+            output_scores           = True,
+            use_cache               = True,
+        )
+
+    # Kumpulkan hasil per trial
+    trials = []
+    for i in range(int(num_trials)):
+        gen_ids  = out.sequences[i][inputs.input_ids.shape[1]:]
+        gen_text = tokenizer.decode(gen_ids, skip_special_tokens=True)
+        scores_i = [s[i:i+1] for s in out.scores]
+        conf     = compute_mtla_confidence(scores_i, gen_ids, K=10)
+        trials.append({
+            "label"    : extract_label(gen_text),
+            "reasoning": extract_reasoning(gen_text),
+            "confidence": conf,
+        })
+
+    # Voting
+    vote              = Counter(t["label"] for t in trials)
+    best_label, count = vote.most_common(1)[0]
+    winners           = [t for t in trials if t["label"] == best_label]
+    avg_conf          = float(np.mean([t["confidence"] for t in winners]))
+    best_reason       = max(winners, key=lambda x: x["confidence"])["reasoning"]
+    is_ambiguous      = count == 1 or avg_conf < 0.45
+
+    emoji, bg, fg, desc = LABEL_INFO.get(best_label, LABEL_INFO["unknown"])
+    label_display = f"{emoji} {best_label.upper().replace('_', ' ')}"
+    conf_pct      = f"{avg_conf * 100:.1f}%"
+    consistency   = f"{count}/{int(num_trials)}"
+    ambig_status  = "⚠️ Ambigu — model ragu-ragu" if is_ambiguous else "✅ Model yakin"
+
+    label_html = f"""
+    <div style="
+        background:{bg}; color:{fg};
+        padding:12px 24px; border-radius:12px;
+        font-size:1.4rem; font-weight:700;
+        text-align:center; display:inline-block;
+        border: 2px solid {fg}30; margin:8px 0;
+    ">
+        {emoji} {best_label.upper().replace('_', ' ')}
+    </div>
+    """
+
+    trial_data = [
+        [
+            f"Trial {i+1}",
+            f"{LABEL_INFO.get(t['label'], LABEL_INFO['unknown'])[0]} "
+            f"{t['label'].upper().replace('_', ' ')}",
+            f"{t['confidence'] * 100:.1f}%",
+            t['reasoning'][:150] + "..." if len(t['reasoning']) > 150 else t['reasoning'],
+        ]
+        for i, t in enumerate(trials)
+    ]
+
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    return (
+        label_display, conf_pct, consistency,
+        ambig_status, desc, best_reason,
+        trial_data, label_html,
+    )
+
+# ================================================================
+# GRADIO UI
+# ================================================================
+css = """
+    .gradio-container { max-width: 900px !important; margin: auto; }
+    footer { display: none !important; }
+"""
+
+with gr.Blocks(
+    title = "DFK Content Classifier",
+    theme = gr.themes.Soft(primary_hue="red", neutral_hue="slate"),
+    css   = css,
+) as demo:
+
+    gr.HTML("""
+    <div style="text-align:center;padding:1.5rem 0 0.5rem">
+        <h1 style="font-size:2rem;font-weight:800;color:#1e293b;margin:0">
+            🛡️ DFK Content Classifier
+        </h1>
+        <p style="color:#64748b;margin:8px 0 4px">
+            Deteksi Disinformasi · Fitnah · Ujaran Kebencian · Fakta
+        </p>
+        <p style="color:#94a3b8;font-size:0.85rem;margin:0">
+            Model: <b>Ministral-3-8B</b> + LoRA Fine-tuning · Bahasa Indonesia
+        </p>
+        <div style="background:#fef9c3;color:#854d0e;padding:6px 16px;border-radius:8px;font-size:0.82rem;margin:8px auto;display:inline-block">
+            ⏳ CPU Mode — estimasi waktu inference: 2-5 menit per request
+        </div>
+        <div style="display:flex;justify-content:center;gap:8px;margin:12px 0;flex-wrap:wrap">
+            <span style="background:#dcfce7;color:#166534;padding:3px 12px;border-radius:20px;font-size:0.82rem">🟢 Fakta</span>
+            <span style="background:#fee2e2;color:#991b1b;padding:3px 12px;border-radius:20px;font-size:0.82rem">🔴 Disinformasi</span>
+            <span style="background:#ffedd5;color:#9a3412;padding:3px 12px;border-radius:20px;font-size:0.82rem">🟠 Fitnah</span>
+            <span style="background:#f1f5f9;color:#1e293b;padding:3px 12px;border-radius:20px;font-size:0.82rem">⚫ Ujaran Kebencian</span>
+        </div>
+    </div>
+    """)
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            text_input = gr.Textbox(
+                label       = "📝 Teks yang ingin diklasifikasi",
+                placeholder = "Masukkan klaim, berita, atau konten...",
+                lines       = 5,
+            )
+            with gr.Row():
+                num_trials = gr.Slider(
+                    minimum = 1, maximum = 5, value = 3, step = 1,
+                    label   = "Jumlah Trial",
+                    info    = "Lebih banyak = lebih akurat tapi lebih lambat",
+                )
+                temperature = gr.Slider(
+                    minimum = 0.1, maximum = 1.0, value = 0.7, step = 0.1,
+                    label   = "Temperature",
+                    info    = "0.1 = deterministik, 1.0 = kreatif",
+                )
+            classify_btn = gr.Button(
+                "🔍 Klasifikasi Sekarang",
+                variant = "primary",
+                size    = "lg",
+            )
+
+        with gr.Column(scale=1):
+            gr.Examples(
+                examples = [
+                    ["Air rebusan bawang putih bisa menyembuhkan virus COVID dalam 24 jam!"],
+                    ["BPOM mengkonfirmasi vaksin COVID-19 sudah melalui uji klinis tiga fase sesuai standar WHO."],
+                    ["Gubernur X terbukti korupsi dana bansos, ada bukti transfer ke rekening keluarganya."],
+                    ["Orang dari suku X itu memang tidak bisa dipercaya dan selalu bikin masalah."],
+                ],
+                inputs = text_input,
+                label  = "💡 Contoh Teks",
+            )
+
+    gr.HTML("<hr style='margin:1rem 0;border-color:#e2e8f0'>")
+
+    label_html_out = gr.HTML()
+
+    with gr.Row():
+        label_out = gr.Textbox(
+            label="🏷️ Label", interactive=False,
+        )
+        conf_out = gr.Textbox(
+            label="🎯 Trust Score (MTLA)", interactive=False,
+            info="Keyakinan model via Multi-Token Logit Averaging",
+        )
+        consistency_out = gr.Textbox(
+            label="🗳️ Konsistensi", interactive=False,
+        )
+        ambig_out = gr.Textbox(
+            label="📊 Status", interactive=False,
+        )
+
+    desc_out = gr.Textbox(
+        label="📖 Deskripsi Label", interactive=False,
+    )
+    reasoning_out = gr.Textbox(
+        label="💬 Reasoning Model", lines=4, interactive=False,
+        info="Penjelasan model tentang keputusannya",
+    )
+
+    with gr.Accordion("🔬 Detail Per Trial", open=False):
+        trial_table = gr.Dataframe(
+            headers = ["Trial", "Label", "Trust Score", "Reasoning"],
+            wrap    = True,
+        )
+
+    with gr.Accordion("🔌 Cara Pakai via API", open=False):
+        gr.Markdown("""
+### Python
+```python
+from gradio_client import Client
+
+client = Client("ggapar/dfk-classifier")
+result = client.predict(
+    text        = "Teks yang ingin dicek",
+    num_trials  = 3,
+    temperature = 0.7,
+    api_name    = "/classify_dfk"
+)
+# result[0] = Label, result[1] = Trust Score, result[5] = Reasoning
+print(result[0], result[1], result[5])
+```
+
+### Install
+```bash
+pip install gradio_client
+```
+        """)
+
+    gr.HTML("""
+    <div style="text-align:center;color:#94a3b8;font-size:0.78rem;margin-top:1rem">
+        Model: <a href="https://huggingface.co/ggapar/Ministral-3-8B-Base-2512-DFK"
+        target="_blank">ggapar/Ministral-3-8B-Base-2512-DFK</a> ·
+        AITF Team 2025
+    </div>
+    """)
+
+    outputs = [
+        label_out, conf_out, consistency_out,
+        ambig_out, desc_out, reasoning_out,
+        trial_table, label_html_out,
+    ]
+    classify_btn.click(fn=classify_dfk,
+                       inputs=[text_input, num_trials, temperature],
+                       outputs=outputs)
+    text_input.submit(fn=classify_dfk,
+                      inputs=[text_input, num_trials, temperature],
+                      outputs=outputs)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+torch>=2.0.0
+transformers>=4.40.0
+peft>=0.10.0
+gradio>=4.0.0
+numpy>=1.24.0
+accelerate>=0.27.0
+gradio_client>=0.6.0