| --- |
| base_model: unsloth/Qwen3.5-0.8B |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - base_model:adapter:unsloth/Qwen3.5-0.8B |
| - lora |
| - sft |
| - transformers |
| - trl |
| - unsloth |
| - dfk-detection |
| - vlm |
| - indonesian |
| - multimodal |
| - image-classification |
| - content-moderation |
| --- |
| |
| <style> |
| .mc { |
| --bg: #f4f0f8; |
| --panel: rgba(255,255,255,0.75); |
| --accent: #8b7bc6; |
| --accent2: #b5a8d9; |
| --accent3: #d4cce8; |
| --peach: #e8a87c; |
| --peach2: #f0c4a0; |
| --border: #d0c8e0; |
| --text: #3a3050; |
| --muted: #7a6f8e; |
| --bright: #5a4a80; |
| --white: #ffffff; |
| --glow: rgba(139,123,198,0.10); |
| --mono: 'JetBrains Mono', monospace; |
| --sans: 'Inter', sans-serif; |
|
|
| font-family: var(--sans); |
| color: var(--text); |
| background: var(--bg); |
| max-width: 940px; |
| margin: 0 auto; |
| padding: 0 0 48px; |
| line-height: 1.7; |
| font-size: 1rem; |
| } |
|
|
| /* ── Hero ── */ |
| .mc-hero { |
| position: relative; |
| border-bottom: 1px solid var(--accent2); |
| margin: 0; |
| } |
| .mc-hero img { |
| display: block; |
| width: 100%; |
| margin: 0; |
| } |
| .mc-title { |
| text-align: center; |
| position: relative; |
| z-index: 1; |
| } |
| .mc-card { |
| display: inline-block; |
| padding: 20px 48px; |
| background: rgba(255,255,255,0.9); |
| border: 1px solid var(--accent2); |
| box-shadow: 0 4px 32px rgba(139,123,198,0.18); |
| margin-top: -70px; |
| } |
| .mc-name { |
| font-size: 1.6rem; |
| font-weight: 900; |
| letter-spacing: 3px; |
| text-transform: uppercase; |
| margin: 0 0 6px; |
| line-height: 1.2; |
| background: linear-gradient(90deg, var(--bright), var(--accent), var(--peach)); |
| background-clip: text; |
| -webkit-background-clip: text; |
| -webkit-text-fill-color: transparent; |
| } |
| .mc-base { |
| font-family: var(--mono); |
| font-size: 0.66rem; |
| color: var(--muted); |
| letter-spacing: 3px; |
| text-transform: uppercase; |
| } |
| |
| /* ── Diamond separators ── */ |
| .mc-sep { |
| display: flex; |
| align-items: center; |
| margin: 0 32px; |
| padding: 16px 0; |
| } |
| .mc-sep-line { |
| flex: 1; |
| height: 1px; |
| background: linear-gradient(90deg, transparent, var(--accent2), transparent); |
| } |
| .mc-dia { |
| width: 9px; |
| height: 9px; |
| background: var(--accent); |
| transform: rotate(45deg); |
| box-shadow: 0 0 8px rgba(139,123,198,0.3); |
| margin: 0 14px; |
| flex-shrink: 0; |
| } |
| |
| /* ── Sections ── */ |
| .mc-section { |
| margin: 0 32px; |
| padding: 28px 32px; |
| background: var(--panel); |
| border: 1px solid var(--border); |
| box-shadow: 0 2px 20px var(--glow); |
| } |
| |
| /* ── Section headers ── */ |
| .mc-shead { |
| text-align: center; |
| margin-bottom: 22px; |
| } |
| .mc-emblem { |
| width: 36px; |
| height: 36px; |
| border: 2px solid var(--accent2); |
| transform: rotate(45deg); |
| margin: 0 auto 12px; |
| display: flex; |
| align-items: center; |
| justify-content: center; |
| box-shadow: 0 0 12px rgba(139,123,198,0.12); |
| background: var(--white); |
| } |
| .mc-glyph { |
| transform: rotate(-45deg); |
| font-size: 0.9rem; |
| color: var(--accent); |
| line-height: 1; |
| } |
| .mc-stitle { |
| font-size: 1.3rem; |
| font-weight: 800; |
| letter-spacing: 4px; |
| text-transform: uppercase; |
| margin: 0 !important; |
| padding: 0 !important; |
| border: none !important; |
| display: block; |
| background: linear-gradient(90deg, var(--accent), var(--bright), var(--accent)); |
| background-clip: text; |
| -webkit-background-clip: text; |
| -webkit-text-fill-color: transparent; |
| } |
| |
| /* ── Body ── */ |
| .mc-sbody p { margin: 0 0 14px; font-size: 0.94rem; } |
| .mc-sbody p:last-child { margin-bottom: 0; } |
| |
| /* ── Sub-headings ── */ |
| .mc-sub { |
| color: var(--bright) !important; |
| font-size: 1rem !important; |
| margin: 22px 0 12px !important; |
| padding: 0 0 7px !important; |
| font-weight: 700; |
| text-transform: uppercase; |
| letter-spacing: 2px; |
| border: none !important; |
| border-bottom: 2px solid var(--accent3) !important; |
| } |
| |
| /* ── Data boxes ── */ |
| .mc-data { |
| background: var(--white); |
| padding: 16px; |
| border: 1px solid var(--border); |
| border-left: 3px solid var(--accent); |
| margin-bottom: 16px; |
| box-shadow: 0 2px 12px var(--glow); |
| font-size: 0.94rem; |
| } |
| .mc-data:last-child { margin-bottom: 0; } |
| .mc-data--peach { border-left-color: var(--peach); } |
| .mc-row { |
| display: flex; |
| align-items: center; |
| margin-bottom: 7px; |
| padding: 5px 0; |
| border-bottom: 1px solid rgba(208,200,224,0.4); |
| } |
| .mc-row:last-child { margin-bottom: 0; border-bottom: none; } |
| .mc-mark { |
| width: 6px; |
| height: 6px; |
| background: var(--accent); |
| transform: rotate(45deg); |
| box-shadow: 0 0 4px rgba(139,123,198,0.25); |
| margin-right: 12px; |
| flex-shrink: 0; |
| } |
| .mc-mark--peach { background: var(--peach); box-shadow: 0 0 4px rgba(232,168,124,0.3); } |
| .mc-label { |
| color: var(--muted); |
| font-weight: 700; |
| margin-right: 12px; |
| min-width: 100px; |
| text-transform: uppercase; |
| letter-spacing: 1px; |
| font-size: 0.85rem; |
| } |
| .mc-val { |
| color: var(--text); |
| font-size: 0.9rem; |
| } |
| |
| /* ── Grid layout ── */ |
| .mc-grid { |
| display: grid; |
| grid-template-columns: 1fr 1fr; |
| gap: 16px; |
| } |
| |
| /* ── Metric cards ── */ |
| .mc-metrics { |
| display: grid; |
| grid-template-columns: repeat(4, 1fr); |
| gap: 12px; |
| margin-bottom: 18px; |
| } |
| .mc-metric { |
| background: var(--white); |
| border: 1px solid var(--border); |
| border-top: 3px solid var(--accent2); |
| padding: 16px 12px 14px; |
| text-align: center; |
| box-shadow: 0 2px 12px var(--glow); |
| } |
| .mc-metric--highlight { border-top-color: var(--peach); } |
| .mc-metric-val { |
| font-family: var(--mono); |
| font-size: 1.7rem; |
| font-weight: 900; |
| color: var(--bright); |
| line-height: 1; |
| margin-bottom: 5px; |
| } |
| .mc-metric--highlight .mc-metric-val { color: var(--peach); } |
| .mc-metric-label { |
| font-family: var(--mono); |
| font-size: 0.56rem; |
| font-weight: 700; |
| letter-spacing: 0.14em; |
| text-transform: uppercase; |
| color: var(--muted); |
| } |
| |
| /* ── Note ── */ |
| .mc-note { |
| border: 1px solid var(--border); |
| border-left: 3px solid var(--peach); |
| padding: 12px 16px; |
| margin-top: 16px; |
| background: rgba(232,168,124,0.06); |
| font-size: 0.86rem; |
| color: var(--muted); |
| } |
| .mc-note strong { |
| color: var(--peach); |
| font-family: var(--mono); |
| font-size: 0.7rem; |
| letter-spacing: 0.1em; |
| text-transform: uppercase; |
| } |
| |
| /* ── Links ── */ |
| .mc a { |
| color: var(--bright); |
| text-decoration: none; |
| font-weight: 600; |
| border-bottom: 1px dotted var(--accent2); |
| } |
| .mc a:hover { color: var(--accent); border-bottom-style: solid; } |
| |
| /* ── Dropdown ── */ |
| .mc-drop { margin-top: 18px; } |
| .mc-drop details { |
| border: 1px solid var(--border); |
| background: var(--white); |
| box-shadow: 0 2px 12px var(--glow); |
| } |
| .mc-drop summary { |
| cursor: pointer; |
| padding: 11px 16px; |
| color: var(--muted); |
| font-size: 0.9rem; |
| font-weight: 700; |
| text-transform: uppercase; |
| letter-spacing: 2px; |
| list-style: none; |
| display: flex; |
| align-items: center; |
| gap: 10px; |
| } |
| .mc-drop summary::-webkit-details-marker { display: none; } |
| .mc-drop summary::before { |
| content: '+'; |
| color: var(--accent); |
| font-size: 1.1rem; |
| font-weight: 700; |
| line-height: 1; |
| flex-shrink: 0; |
| } |
| .mc-drop details[open] summary::before { content: '−'; } |
| .mc-drop summary:hover { color: var(--bright); } |
| .mc-drop-body { |
| padding: 18px; |
| border-top: 1px solid var(--border); |
| background: rgba(244,240,248,0.4); |
| } |
| .mc-drop-body p { margin: 0 0 12px; font-size: 0.9rem; } |
| |
| /* ── Code ── */ |
| .mc pre { |
| background: #2a2240; |
| padding: 14px 16px; |
| margin: 0 0 18px; |
| border: 1px solid var(--accent3); |
| border-left: 3px solid var(--accent); |
| overflow-x: auto; |
| color: #d4cce8; |
| box-shadow: 0 2px 12px var(--glow); |
| } |
| .mc pre:last-child { margin-bottom: 0; } |
| .mc pre code { |
| font-family: var(--mono); |
| font-size: 0.74rem; |
| line-height: 1.6; |
| background: none; |
| color: inherit; |
| padding: 0; |
| display: block; |
| border: none; |
| } |
| .mc code { |
| font-family: var(--mono); |
| font-size: 0.85em; |
| color: var(--bright); |
| background: rgba(139,123,198,0.08); |
| padding: 2px 6px; |
| border: 1px solid rgba(139,123,198,0.12); |
| } |
| </style> |
| <html lang="en"> |
| <head> |
| <meta charset="UTF-8"> |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> |
| <title>KomdigiITS-0.8B-DFK-MultimodalClassification</title> |
| <link rel="preconnect" href="https://fonts.googleapis.com"> |
| <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin> |
| <link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;600;700;800;900&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet"> |
| </head> |
| <body> |
| <div class="mc"> |
| |
| <div class="mc-hero"> |
| <img src="dfk_hero_banner.png" alt="image"> |
| </div> |
| <div class="mc-title"> |
| <div class="mc-card"> |
| <h1 class="mc-name">KomdigiITS-0.8B-DFK<br>Multimodal Classification</h1> |
| <span class="mc-base">Qwen3.5-0.8B · LoRA · Vision-Language</span> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Overview</span> |
| </div> |
| <div class="mc-sbody"> |
| <p>A LoRA adapter fine-tuned on <a href="https://huggingface.co/unsloth/Qwen3.5-0.8B">Qwen3.5-0.8B</a> as a Vision-Language Model for multimodal content classification. The model analyzes social media screenshots and classifies them into four categories: <code>netral</code>, <code>disinformasi</code>, <code>fitnah</code>, and <code>ujaran kebencian</code>.</p> |
| <p>Trained using the <a href="https://github.com/aitf-its-tim3-dfk/SITA">SITA</a> framework with Unsloth's SFT pipeline. Given an image, the model produces a structured analysis with a classification label and a detailed Indonesian-language reasoning of any violations found.</p> |
| <div class="mc-note"> |
| <strong>♦ Note:</strong> This is the final checkpoint from Workshop 3 (<code>final-qwen35-0.8b-ws3</code>), trained on the DFK VLM Dataset V3 with augmented train/val splits. |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Model Details</span> |
| </div> |
| <div class="mc-sbody"> |
| <div class="mc-grid"> |
| <div> |
| <h3 class="mc-sub">Identity</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Developed by:</span><span class="mc-val">DFK Tim 3 ITS</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Model type:</span><span class="mc-val">VLM — LoRA adapter</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Language:</span><span class="mc-val">Indonesian</span></div> |
| </div> |
| </div> |
| <div> |
| <h3 class="mc-sub">Architecture</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Base model:</span><span class="mc-val"><a href="https://huggingface.co/unsloth/Qwen3.5-0.8B">unsloth/Qwen3.5-0.8B</a></span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Arch:</span><span class="mc-val">Qwen3_5ForCausalLM</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Parameters:</span><span class="mc-val">0.8B (base)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Precision:</span><span class="mc-val">bfloat16</span></div> |
| </div> |
| </div> |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Uses</span> |
| </div> |
| <div class="mc-sbody"> |
| <h3 class="mc-sub">Direct Use</h3> |
| <div class="mc-data"> |
| <p style="margin:0;">Image-based content moderation classification for Indonesian social media. Given a screenshot, the model produces a structured analysis with a classification label (<code>netral</code>, <code>disinformasi</code>, <code>fitnah</code>, or <code>ujaran kebencian</code>) and a detailed reasoning in Indonesian.</p> |
| </div> |
| <h3 class="mc-sub">Out-of-Scope Use</h3> |
| <div class="mc-data mc-data--peach"> |
| <p style="margin:0;">This model is not intended for general-purpose vision-language tasks. It is specialized for the DFK disinformation detection pipeline and should not be used for content moderation in other languages or domains without further fine-tuning.</p> |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Evaluation</span> |
| </div> |
| <div class="mc-sbody"> |
| <p>Evaluated on the held-out validation split using greedy decoding (<code>temperature=0.0</code>) and BERTScore (<code>bert-base-multilingual-cased</code>).</p> |
| <div class="mc-metrics"> |
| <div class="mc-metric mc-metric--highlight"> |
| <div class="mc-metric-val">92.5</div> |
| <div class="mc-metric-label">Accuracy</div> |
| </div> |
| <div class="mc-metric"> |
| <div class="mc-metric-val">89.3</div> |
| <div class="mc-metric-label">F1 Macro</div> |
| </div> |
| <div class="mc-metric"> |
| <div class="mc-metric-val">92.8</div> |
| <div class="mc-metric-label">F1 Weighted</div> |
| </div> |
| <div class="mc-metric"> |
| <div class="mc-metric-val">79.5</div> |
| <div class="mc-metric-label">BERTScore F1</div> |
| </div> |
| </div> |
| <div class="mc-drop"> |
| <details> |
| <summary>Per-Class Breakdown</summary> |
| <div class="mc-drop-body"> |
| <div class="mc-data" style="margin-bottom:0;"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Netral:</span><span class="mc-val">P 0.954 · R 0.941 · F1 0.948 · n=970</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Ujaran Kbnci:</span><span class="mc-val">P 0.982 · R 0.930 · F1 0.955 · n=867</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Disinformasi:</span><span class="mc-val">P 0.943 · R 0.888 · F1 0.915 · n=392</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Fitnah:</span><span class="mc-val">P 0.651 · R 0.901 · F1 0.756 · n=213</span></div> |
| </div> |
| </div> |
| </details> |
| </div> |
| <div class="mc-drop"> |
| <details> |
| <summary>BERTScore Details</summary> |
| <div class="mc-drop-body"> |
| <div class="mc-data" style="margin-bottom:0;"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Precision:</span><span class="mc-val">0.797</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Recall:</span><span class="mc-val">0.793</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">F1:</span><span class="mc-val">0.795</span></div> |
| </div> |
| </div> |
| </details> |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Training Details</span> |
| </div> |
| <div class="mc-sbody"> |
| <h3 class="mc-sub">Training Data</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Dataset:</span><span class="mc-val"><code>dfk_vlm_dataset_v3</code> (augmented on <code>fitnah</code> class)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Split mode:</span><span class="mc-val">Fixed splits (train_aug.csv / val_aug.csv)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Train size:</span><span class="mc-val">14,293 samples</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Val size:</span><span class="mc-val">2,831 samples</span></div> |
| </div> |
| <h3 class="mc-sub">Label Classes</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Netral:</span><span class="mc-val">Factual content or non-DFK material — no violation detected</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Disinformasi:</span><span class="mc-val">Claims that contradict established facts, not directed at a specific person</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Fitnah:</span><span class="mc-val">False claims directed at a specific individual (defamation)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Ujaran Kbnci:</span><span class="mc-val">Hate speech targeting ethnicity, religion, race, or intergroup identity (SARA)</span></div> |
| </div> |
| <div class="mc-drop"> |
| <details> |
| <summary>Dataset Distribution</summary> |
| <div class="mc-drop-body"> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label" style="min-width:140px;">Train (aug)</span><span class="mc-val" style="font-family:var(--mono);font-size:0.82rem;">14,293 total</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Netral:</span><span class="mc-val">3,883 (27.2%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Fitnah:</span><span class="mc-val">3,846 (26.9%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Ujaran Kbnci:</span><span class="mc-val">3,484 (24.4%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Disinformasi:</span><span class="mc-val">3,080 (21.6%)</span></div> |
| </div> |
| <div class="mc-data" style="margin-top:12px;"> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label" style="min-width:140px;">Val (aug)</span><span class="mc-val" style="font-family:var(--mono);font-size:0.82rem;">2,831 total</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Netral:</span><span class="mc-val">970 (34.3%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Ujaran Kbnci:</span><span class="mc-val">867 (30.6%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Disinformasi:</span><span class="mc-val">765 (27.0%)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label" style="min-width:140px;">Fitnah:</span><span class="mc-val">229 (8.1%)</span></div> |
| </div> |
| </div> |
| </details> |
| </div> |
| <div class="mc-grid"> |
| <div> |
| <h3 class="mc-sub">LoRA Configuration</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">r:</span><span class="mc-val">16</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Alpha:</span><span class="mc-val">16</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Dropout:</span><span class="mc-val">0.1</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Targets:</span><span class="mc-val">all-linear</span></div> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">Vision:</span><span class="mc-val">✓ finetuned</span></div> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">Language:</span><span class="mc-val">✓ finetuned</span></div> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">Attention:</span><span class="mc-val">✓ finetuned</span></div> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">MLP:</span><span class="mc-val">✓ finetuned</span></div> |
| </div> |
| </div> |
| <div> |
| <h3 class="mc-sub">Hyperparameters</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Epochs:</span><span class="mc-val">3</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Batch size:</span><span class="mc-val">32</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">LR:</span><span class="mc-val">2e-4</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Optimizer:</span><span class="mc-val">AdamW 8-bit</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Max seq len:</span><span class="mc-val">2048</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Grad accum:</span><span class="mc-val">1</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Grad ckpt:</span><span class="mc-val">unsloth</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Seed:</span><span class="mc-val">3407</span></div> |
| </div> |
| </div> |
| </div> |
| <h3 class="mc-sub">Trainer</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Type:</span><span class="mc-val"><code>unsloth_vlm_sft</code> (Unsloth VLM SFT trainer)</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Train on:</span><span class="mc-val">Responses only</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Instr part:</span><span class="mc-val"><code><|im_start|>user\n</code></span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Resp part:</span><span class="mc-val"><code><|im_start|>assistant\n</code></span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Best model:</span><span class="mc-val">Selected by <code>eval_loss</code> (lower is better)</span></div> |
| </div> |
| <div class="mc-drop"> |
| <details> |
| <summary>Prompt Template</summary> |
| <div class="mc-drop-body"> |
| <p>Each sample is formatted as a multi-turn conversation using <code>qwen3.5_chatml</code>:</p> |
| <pre><code><|im_start|>user |
| Anda adalah seorang analis konten media sosial ahli. Diberikan tangkapan layar |
| dari sebuah konten, tentukan label kategori pelanggaran dan berikan analisis |
| detail mengenai pelanggaran yang ditemukan. |
| Ringkasan: {ringkasan} |
| Klaim: {klaim} |
| Fakta: {fakta} |
| <image> |
| <|im_end|> |
| <|im_start|>assistant |
| Label: {label} |
| |
| Analisis: {analisis} |
| <|im_end|></code></pre> |
| <h3 class="mc-sub" style="margin-top:20px;">Input Fields</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Ringkasan:</span><span class="mc-val">Content summary. In the RAG pipeline this is the concatenation of the image caption (from a captioning model) and any user-provided text (e.g. post caption, tweet text). Effectively holds all available textual context about the content.</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Klaim:</span><span class="mc-val">The core claim extracted from the content, used as a web search query for fact-checking. Generated by an LLM from the ringkasan. Can also be a direct caption or user-provided text in simpler setups.</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Fakta:</span><span class="mc-val">Verification context retrieved via web search. Contains numbered search results with titles, descriptions, and source URLs. If no relevant sources are found, defaults to <code>"Tidak ditemukan sumber yang valid."</code></span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label"><image>:</span><span class="mc-val">Screenshot of the social media post being analyzed.</span></div> |
| </div> |
| <h3 class="mc-sub">Output Fields</h3> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">Label:</span><span class="mc-val">One of <code>netral</code>, <code>disinformasi</code>, <code>fitnah</code>, or <code>ujaran kebencian</code>.</span></div> |
| <div class="mc-row"><span class="mc-mark mc-mark--peach"></span><span class="mc-label">Analisis:</span><span class="mc-val">Free-form Indonesian-language explanation of why the content was assigned its label, referencing the image, context, and any retrieved facts.</span></div> |
| </div> |
| </div> |
| </details> |
| </div> |
| |
| <div class="mc-drop"> |
| <details> |
| <summary>Full Training Config</summary> |
| <div class="mc-drop-body"> |
| <pre><code>experiment_name: final-qwen35-0.8b-ws3 |
| seed: 3407 |
| |
| reporting: |
| wandb: true |
| wandb_project: "DFK3" |
| |
| model: |
| name: unsloth_vlm |
| pretrained: unsloth/Qwen3.5-0.8B |
| kwargs: |
| load_in_4bit: false |
| chat_template: "sita/templates/qwen3.5_chatml.jinja" |
| |
| adapter: |
| name: unsloth_vlm_lora |
| kwargs: |
| finetune_vision_layers: true |
| finetune_language_layers: true |
| finetune_attention_modules: true |
| finetune_mlp_modules: true |
| r: 16 |
| lora_alpha: 16 |
| lora_dropout: 0.1 |
| bias: "none" |
| target_modules: "all-linear" |
| use_gradient_checkpointing: "unsloth" |
| random_state: 3407 |
| |
| dataset: |
| name: dfk_vlm_dataset_v3 |
| |
| training: |
| num_epochs: 3 |
| batch_size: 32 |
| learning_rate: 2e-4 |
| gradient_accumulation_steps: 1 |
| logging_steps: 1 |
| save_steps: 100 |
| eval_steps: 50 |
| extra: |
| seed: 3407 |
| max_length: 2048 |
| load_best_model_at_end: true |
| metric_for_best_model: eval_loss |
| greater_is_better: false |
| |
| trainer: |
| name: unsloth_vlm_sft |
| kwargs: |
| train_on_responses_only: true |
| instruction_part: "<|im_start|>user\n" |
| response_part: "<|im_start|>assistant\n" |
| optim: adamw_8bit |
| |
| evaluation: |
| name: vlm_gen |
| kwargs: |
| max_new_tokens: 512 |
| temperature: 0.0 |
| bert_model: bert-base-multilingual-cased |
| batch_size: 16 |
| num_workers: 11</code></pre> |
| |
| </div> |
| </details> |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Model Sources</span> |
| </div> |
| <div class="mc-sbody"> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Framework:</span><span class="mc-val"><a href="https://github.com/aitf-its-tim3-dfk/SITA">SITA</a></span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">W&B Run:</span><span class="mc-val"><a href="https://wandb.ai/aitfits2026-kementerian-komdigi/DFK3/runs/9ygdkx50">DFK3 / final-qwen35-0.8b-ws3</a></span></div> |
| </div> |
| </div> |
| </div> |
| |
| <div class="mc-sep"><div class="mc-sep-line"></div><div class="mc-dia"></div><div class="mc-sep-line"></div></div> |
| |
| <div class="mc-section"> |
| <div class="mc-shead"> |
| <div class="mc-emblem"><span class="mc-glyph">✺</span></div> |
| <span class="mc-stitle">Framework Versions</span> |
| </div> |
| <div class="mc-sbody"> |
| <div class="mc-data"> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">TRL:</span><span class="mc-val">0.22.2</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Transformers:</span><span class="mc-val">5.3.0</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">PyTorch:</span><span class="mc-val">2.11.0+cu128</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Datasets:</span><span class="mc-val">4.3.0</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">PEFT:</span><span class="mc-val">0.19.0</span></div> |
| <div class="mc-row"><span class="mc-mark"></span><span class="mc-label">Tokenizers:</span><span class="mc-val">0.22.2</span></div> |
| </div> |
| </div> |
| </div> |
| |
| </div> |
| </body> |
| </html> |
| |
