Instructions to use aitf-its-tim3-dfk/ministral-8b-merged-ws3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use aitf-its-tim3-dfk/ministral-8b-merged-ws3 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="aitf-its-tim3-dfk/ministral-8b-merged-ws3")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("aitf-its-tim3-dfk/ministral-8b-merged-ws3")
model = AutoModelForMultimodalLM.from_pretrained("aitf-its-tim3-dfk/ministral-8b-merged-ws3")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use aitf-its-tim3-dfk/ministral-8b-merged-ws3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aitf-its-tim3-dfk/ministral-8b-merged-ws3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-its-tim3-dfk/ministral-8b-merged-ws3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aitf-its-tim3-dfk/ministral-8b-merged-ws3

SGLang

How to use aitf-its-tim3-dfk/ministral-8b-merged-ws3 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "aitf-its-tim3-dfk/ministral-8b-merged-ws3" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-its-tim3-dfk/ministral-8b-merged-ws3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "aitf-its-tim3-dfk/ministral-8b-merged-ws3" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-its-tim3-dfk/ministral-8b-merged-ws3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use aitf-its-tim3-dfk/ministral-8b-merged-ws3 with Docker Model Runner:
```
docker model run hf.co/aitf-its-tim3-dfk/ministral-8b-merged-ws3
```

ministral-8b-merged-ws3 / README.md

spuuntries

docs: tidy up metrics

521a01e 29 days ago

preview code

Raw

History Blame Contribute Delete

27.6 kB

	---
	base_model:
	- aitf-komdigi/KomdigiITS-8B-DFK-MultimodalClassification
	- aitf-komdigi/KomdigiITS-8B-DFK-TextClassification
	library_name: transformers
	pipeline_tag: text-generation
	tags:
	- merge
	- slerp
	- dfk-detection
	- vlm
	- text-classification
	- indonesian
	- multimodal
	- image-classification
	- content-moderation
	- mistral3
	---

	<style>
	@import url('https://fonts.googleapis.com/css2?family=Playfair+Display:wght@400;600;700;800;900&family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap');

	.fo {
	--bg: #faf8f5;
	--surface: #ffffff;
	--card: #ffffff;
	--tint: #f5f0eb;
	--edge: #e8e0d6;
	--rule: #d8cec4;
	--text: #5c5060;
	--dim: #998da0;
	--bright: #3a2e40;
	--cream: #f8f4ef;
	--warm: #c48a6a;
	--warm2: #daa88a;
	--rose: #c8909c;
	--rose2: #e0b0ba;
	--sage: #7aac96;
	--sage2: #9cc8b4;
	--w-glow: rgba(196,138,106,0.06);
	--s-glow: rgba(122,172,150,0.06);
	--r-glow: rgba(200,144,156,0.05);
	--shadow: rgba(90,70,80,0.06);
	--mono: 'JetBrains Mono', monospace;
	--sans: 'Inter', sans-serif;
	--serif: 'Playfair Display', Georgia, serif;

	font-family: var(--sans);
	color: var(--text);
	max-width: 860px;
	margin: 0 auto;
	padding: 0 0 64px;
	line-height: 1.75;
	font-size: 0.95rem;
	background:
	radial-gradient(ellipse at 30% 0%, rgba(196,138,106,0.04) 0%, transparent 50%),
	radial-gradient(ellipse at 70% 0%, rgba(122,172,150,0.04) 0%, transparent 50%),
	var(--bg);
	}

	/* ── Hero ── */
	.fo-hero { position: relative; overflow: hidden; border-radius: 0 0 12px 12px; }
	.fo-hero img {
	display: block; width: 100%; height: 300px;
	object-fit: cover; margin: 0;
	}
	.fo-ident {
	position: absolute; bottom: 0; left: 0; right: 0;
	padding: 90px 44px 32px;
	background: linear-gradient(to top,
	rgba(58,46,64,0.88) 0%,
	rgba(58,46,64,0.6) 40%,
	transparent 100%);
	}
	.fo-name {
	font-family: var(--serif);
	font-size: 2.6rem; font-weight: 800;
	color: #fff; letter-spacing: 0.01em;
	line-height: 1.08; margin: 0 0 8px;
	text-shadow: 0 2px 16px rgba(0,0,0,0.25);
	}
	.fo-tagline {
	font-family: var(--mono); font-size: 0.58rem;
	color: var(--rose2); letter-spacing: 0.16em;
	text-transform: uppercase; display: block;
	}
	.fo-badge {
	display: inline-flex; align-items: center; gap: 7px;
	margin-top: 12px; padding: 4px 12px;
	border: 1px solid rgba(255,255,255,0.2);
	border-radius: 20px;
	background: rgba(122,172,150,0.15);
	font-family: var(--mono); font-size: 0.52rem; font-weight: 600;
	letter-spacing: 0.12em; text-transform: uppercase; color: var(--sage2);
	backdrop-filter: blur(4px);
	}
	.fo-badge::before {
	content: '\2726'; font-size: 0.5rem; color: var(--rose2); opacity: 0.8;
	}

	/* ── Section ── */
	.fo-section { padding: 0; }
	.fo-shead {
	display: flex; align-items: baseline; gap: 14px;
	padding: 24px 44px 0; margin-bottom: 24px;
	}
	.fo-snum {
	font-family: var(--mono); font-size: 0.58rem; font-weight: 600;
	color: var(--rose); letter-spacing: 0.08em;
	flex-shrink: 0; opacity: 0.7;
	}
	.fo-stitle {
	font-family: var(--serif); font-size: 1.35rem; font-weight: 700;
	color: var(--bright); letter-spacing: 0.01em;
	}
	.fo-sbody { padding: 0 44px 40px; }
	.fo-sbody p { margin: 0 0 14px; font-size: 0.93rem; }
	.fo-sbody p:last-child { margin-bottom: 0; }
	.fo strong { color: inherit !important; font-weight: 700 !important; }

	/* ── Sub-headings ── */
	.fo-sub {
	font-family: var(--serif) !important;
	color: var(--bright) !important; font-size: 1.02rem !important;
	margin: 22px 0 12px !important; padding: 0 0 6px !important;
	font-weight: 700; letter-spacing: 0.01em;
	border: none !important;
	border-bottom: 1px solid var(--edge) !important;
	}

	/* ── Cards ── */
	.fo-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; }
	.fo-card {
	background: var(--card); border: 1px solid var(--edge);
	border-radius: 8px; position: relative;
	box-shadow: 0 2px 12px var(--shadow);
	overflow: hidden;
	}
	.fo-card::before {
	content: ''; position: absolute; top: 0; left: 0; right: 0;
	height: 3px; background: linear-gradient(90deg, var(--warm), var(--rose));
	}
	.fo-card--sage::before { background: linear-gradient(90deg, var(--sage), var(--sage2)); }
	.fo-card--muted::before { background: var(--edge); height: 2px; }
	.fo-card--warn::before { background: linear-gradient(90deg, #c47a6a, #d49080); }
	.fo-chead {
	font-family: var(--mono); font-size: 0.58rem; font-weight: 600;
	letter-spacing: 0.12em; text-transform: uppercase; color: var(--dim);
	padding: 10px 16px 8px; border-bottom: 1px solid var(--edge);
	background: var(--tint);
	}
	.fo-chead::after {
	content: ' \2726'; font-size: 0.45rem; color: var(--rose); opacity: 0.45;
	}
	.fo-card--sage .fo-chead::after { color: var(--sage); }
	.fo-row {
	display: grid; grid-template-columns: 12ch 1fr; align-items: baseline;
	column-gap: 8px; padding: 7px 16px;
	border-bottom: 1px solid rgba(232,224,214,0.5); font-size: 0.85rem;
	}
	.fo-row:last-child { border-bottom: none; }
	.fo-k {
	font-family: var(--mono); font-size: 0.78rem; color: var(--dim);
	font-weight: 500;
	}
	.fo-k::after { content: ':'; }
	.fo-v { color: var(--bright); font-size: 0.85rem; }
	.fo-row .fo-v:only-child { grid-column: 1 / -1; }

	/* ── Metric cards ── */
	.fo-metrics { display: grid; grid-template-columns: repeat(3, 1fr); gap: 12px; margin-bottom: 18px; }
	.fo-metric {
	background: var(--card); border: 1px solid var(--edge);
	border-radius: 8px; padding: 16px 10px 14px;
	text-align: center; position: relative;
	box-shadow: 0 2px 12px var(--shadow);
	overflow: hidden;
	}
	.fo-metric::before {
	content: ''; position: absolute; top: 0; left: 0; right: 0;
	height: 3px; background: linear-gradient(90deg, var(--warm), var(--rose));
	}
	.fo-metric--sage::before { background: linear-gradient(90deg, var(--sage), var(--sage2)); }
	.fo-mval {
	font-family: var(--serif); font-size: 1.7rem; font-weight: 800;
	color: var(--bright); line-height: 1; margin-bottom: 5px;
	}
	.fo-metric--sage .fo-mval { color: var(--sage); }
	.fo-mlbl {
	font-family: var(--mono); font-size: 0.5rem; font-weight: 600;
	letter-spacing: 0.12em; text-transform: uppercase; color: var(--dim);
	}

	/* ── Note ── */
	.fo-note {
	border: 1px solid var(--edge); border-radius: 8px;
	padding: 14px 18px; margin-top: 18px;
	background: var(--tint); font-size: 0.85rem; color: var(--dim);
	border-left: 3px solid var(--warm);
	}
	.fo-note strong {
	color: var(--warm) !important; font-family: var(--mono) !important;
	font-size: 0.6rem !important; letter-spacing: 0.1em !important; text-transform: uppercase !important;
	}
	.fo-note--sage { border-left-color: var(--sage); }
	.fo-note--sage strong { color: var(--sage) !important; }

	/* ── Comparison table ── */
	.fo-cmp {
	width: 100%; border-collapse: collapse;
	font-size: 0.76rem; margin-bottom: 18px;
	}
	.fo-cmp th {
	text-align: left; padding: 9px 10px;
	font-family: var(--mono); font-weight: 600;
	letter-spacing: 0.1em; text-transform: uppercase;
	color: var(--dim); border-bottom: 2px solid var(--edge);
	font-size: 0.56rem;
	}
	.fo-cmp td {
	padding: 8px 10px; border-bottom: 1px solid var(--edge);
	color: var(--text); font-family: var(--mono); font-size: 0.76rem;
	}
	.fo-cmp tr:last-child td { border-bottom: none; }
	.fo-cmp-warm { color: var(--warm) !important; font-weight: 600; }
	.fo-cmp-sage { color: var(--sage) !important; font-weight: 700; }
	.fo-cmp-dim { color: var(--dim) !important; }
	.fo-cmp-row-merge td { background: rgba(122,172,150,0.06); }

	/* ── Links ── */
	.fo a { color: var(--bright); text-decoration: none; border-bottom: 1px solid var(--rule); }
	.fo a:hover { color: var(--warm); border-bottom-color: var(--warm); }

	/* ── Dropdown ── */
	.fo details {
	border: 1px solid var(--edge); border-radius: 8px;
	margin-top: 18px; background: var(--card);
	box-shadow: 0 2px 12px var(--shadow);
	overflow: hidden;
	}
	.fo summary {
	list-style: none; padding: 11px 16px; cursor: pointer;
	font-family: var(--mono); font-size: 0.64rem; font-weight: 600;
	letter-spacing: 0.1em; text-transform: uppercase; color: var(--dim);
	user-select: none; display: flex; align-items: center; gap: 8px;
	background: var(--tint);
	}
	.fo summary::-webkit-details-marker { display: none; }
	.fo summary::before {
	content: '+'; color: var(--rose); font-size: 0.85rem; line-height: 1; flex-shrink: 0;
	}
	.fo details[open] summary::before { content: '\2212'; }
	.fo summary:hover { color: var(--bright); }
	.fo details.fo-det--sage summary::before { color: var(--sage); }
	.fo-drop { padding: 18px 16px; border-top: 1px solid var(--edge); }
	.fo-drop p { margin: 0 0 14px; font-size: 0.88rem; }

	/* ── Code ── */
	.fo pre {
	background: #2e2834; border: 1px solid var(--edge);
	border-radius: 6px;
	padding: 14px 16px; overflow-x: auto;
	font-family: var(--mono); font-size: 0.71rem;
	line-height: 1.65; color: #d8d0e0; margin: 0 0 18px;
	}
	.fo pre:last-child { margin-bottom: 0; }
	.fo pre code { background: none !important; color: inherit !important; padding: 0 !important; border: none !important; }
	.fo code {
	font-family: var(--mono) !important; font-size: 0.82em !important; color: var(--warm) !important;
	background: var(--w-glow) !important; padding: 2px 5px !important; border-radius: 3px !important;
	border: 1px solid rgba(196,138,106,0.12) !important;
	}

	/* ── Separator ── */
	.fo-sep {
	display: flex; align-items: center; justify-content: center;
	padding: 6px 44px; color: var(--rose); font-size: 0.55rem;
	letter-spacing: 0.4em; opacity: 0.4;
	}
	.fo-sep::before, .fo-sep::after {
	content: ''; flex: 1; height: 1px;
	background: linear-gradient(90deg, transparent, var(--edge), transparent);
	}
	.fo-sep::before { margin-right: 14px; }
	.fo-sep::after { margin-left: 14px; }

	/* ── Section divider ── */
	.fo-div {
	height: 1px; margin: 0 44px;
	background: linear-gradient(90deg, transparent, var(--edge), transparent);
	}

	/* ── Mobile ── */
	@media (max-width: 640px) {
	.fo-hero img { height: 200px; }
	.fo-name { font-size: 1.9rem; }
	.fo-ident { padding: 70px 24px 24px; }
	.fo-shead { padding: 20px 24px 0; }
	.fo-sbody { padding: 0 24px 32px; }
	.fo-grid { grid-template-columns: 1fr; }
	.fo-metrics { grid-template-columns: repeat(2, 1fr); }
	.fo-row { grid-template-columns: 1fr; gap: 2px; }
	.fo-k::after { content: ''; }
	.fo-sep, .fo-div { margin-left: 24px; margin-right: 24px; padding-left: 24px; padding-right: 24px; }
	}
	</style>
	<html lang="en">
	<head>
	<meta charset="UTF-8">
	<meta name="viewport" content="width=device-width, initial-scale=1.0">
	<title>KomdigiITS-8B-DFK-MergedClassification</title>
	</head>
	<body>
	<div class="fo">

	<div class="fo-hero">
	<img src="dfk_hero_banner.png" alt="image">
	<div class="fo-ident">
	<h1 class="fo-name">KomdigiITS-8B-DFK<br>Merged Classification</h1>
	<span class="fo-tagline">Ministral-3-8B · SLERP Merge · Multimodal + Text</span>
	<div class="fo-badge">SLERP · 0.30 multimodal · 0.70 text</div>
	</div>
	</div>

	<div class="fo-sep">✶</div>

	<!-- ─── 01 · Overview ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">01</span>
	<span class="fo-stitle">Overview</span>
	</div>
	<div class="fo-sbody">
	<p>A <strong>SLERP-merged</strong> full model combining two specialized LoRA adapters into a single unified checkpoint. The merge blends the <a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-MultimodalClassification">8B Multimodal Multimodal classifier</a> with the <a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-TextClassification">text-only DFK classifier</a>, both built on <a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-CPT">KomdigiITS-8B-DFK-CPT</a> (Ministral-3-8B-Base-2512).</p>
	<p>The result is a model that operates in <strong>two modes</strong>: image-based multimodal classification and text-only classification — strong performance on both tasks from a single set of weights.</p>
	<div class="fo-note">
	<strong>✶ Note:</strong> Best SLERP blend from a parameter sweep (<code>w=0.30</code> multimodal, <code>w=0.70</code> text). Parent adapters were merged into base model weights, producing a standalone model with no adapter overhead.
	</div>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 02 · Model Details ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">02</span>
	<span class="fo-stitle">Model Details</span>
	</div>
	<div class="fo-sbody">
	<div class="fo-grid">
	<div class="fo-card">
	<div class="fo-chead">Identity</div>
	<div class="fo-row"><span class="fo-k">Developed</span><span class="fo-v">DFK Tim 1 & Tim 3 ITS</span></div>
	<div class="fo-row"><span class="fo-k">Type</span><span class="fo-v">Merged model (SLERP)</span></div>
	<div class="fo-row"><span class="fo-k">Language</span><span class="fo-v">Indonesian</span></div>
	<div class="fo-row"><span class="fo-k">Modes</span><span class="fo-v">Multimodal + Text-only</span></div>
	</div>
	<div class="fo-card fo-card--sage">
	<div class="fo-chead">Architecture</div>
	<div class="fo-row"><span class="fo-k">Base</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-CPT">KomdigiITS-8B-DFK-CPT</a></span></div>
	<div class="fo-row"><span class="fo-k">Arch</span><span class="fo-v">Mistral3ForConditionalGeneration</span></div>
	<div class="fo-row"><span class="fo-k">Params</span><span class="fo-v">8B (full merged weights)</span></div>
	<div class="fo-row"><span class="fo-k">Precision</span><span class="fo-v">float16</span></div>
	</div>
	</div>
	<h3 class="fo-sub">Parent Models</h3>
	<div class="fo-grid">
	<div class="fo-card">
	<div class="fo-chead">Multimodal Parent · w=0.30</div>
	<div class="fo-row"><span class="fo-k">Model</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-MultimodalClassification">8B-DFK-MultimodalClassification</a></span></div>
	<div class="fo-row"><span class="fo-k">Task</span><span class="fo-v">Multimodal — image + text</span></div>
	<div class="fo-row"><span class="fo-k">Dataset</span><span class="fo-v"><code>dfk_vlm_dataset_v3</code></span></div>
	</div>
	<div class="fo-card fo-card--sage">
	<div class="fo-chead">Text Parent · w=0.70</div>
	<div class="fo-row"><span class="fo-k">Model</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-TextClassification">8B-DFK-TextClassification</a></span></div>
	<div class="fo-row"><span class="fo-k">Task</span><span class="fo-v">Text-only DFK classification</span></div>
	<div class="fo-row"><span class="fo-k">Dataset</span><span class="fo-v"><code>dfk_text_dataset</code></span></div>
	</div>
	</div>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 03 · Uses ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">03</span>
	<span class="fo-stitle">Uses</span>
	</div>
	<div class="fo-sbody">
	<h3 class="fo-sub">Direct Use</h3>
	<div class="fo-card fo-card--muted" style="margin-bottom:14px;">
	<div class="fo-row"><span class="fo-v">This model supports <strong>two input modes</strong> from a single checkpoint:</span></div>
	</div>
	<div class="fo-grid">
	<div class="fo-card">
	<div class="fo-chead">Mode 1 · Multimodal</div>
	<div class="fo-row"><span class="fo-v">Image-based content moderation. Given a social media screenshot with contextual metadata, classifies into 4 labels: <code>netral</code>, <code>disinformasi</code>, <code>fitnah</code>, or <code>ujaran kebencian</code>.</span></div>
	</div>
	<div class="fo-card fo-card--sage">
	<div class="fo-chead">Mode 2 · Text-Only</div>
	<div class="fo-row"><span class="fo-v">Text-based DFK detection using article references. Classifies into 5 labels: <code>Fakta</code>, <code>Disinformasi</code>, <code>Fitnah</code>, <code>Ujaran Kebencian</code>, or <code>Non-DFK</code>.</span></div>
	</div>
	</div>
	<h3 class="fo-sub">Out-of-Scope Use</h3>
	<div class="fo-card fo-card--warn">
	<div class="fo-row"><span class="fo-v">Not intended for general-purpose vision-language or text generation tasks. Specialized for the DFK detection pipeline — should not be used for content moderation in other languages or domains without further fine-tuning.</span></div>
	</div>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 04 · Evaluation ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">04</span>
	<span class="fo-stitle">Evaluation</span>
	</div>
	<div class="fo-sbody">
	<p>Evaluated on held-out validation splits using greedy decoding (<code>temperature=0.0</code>) and BERTScore (<code>bert-base-multilingual-cased</code>).</p>

	<h3 class="fo-sub">Multimodal Task (Image + Text)</h3>
	<div class="fo-metrics">
	<div class="fo-metric fo-metric--sage">
	<div class="fo-mval">88.5</div>
	<div class="fo-mlbl">Accuracy</div>
	</div>
	<div class="fo-metric">
	<div class="fo-mval">89.4</div>
	<div class="fo-mlbl">F1 Weighted</div>
	</div>
	<div class="fo-metric">
	<div class="fo-mval">77.3</div>
	<div class="fo-mlbl">BERTScore F1</div>
	</div>
	</div>

	<h3 class="fo-sub">Text-Only Task</h3>
	<div class="fo-metrics">
	<div class="fo-metric fo-metric--sage">
	<div class="fo-mval">91.0</div>
	<div class="fo-mlbl">Accuracy</div>
	</div>
	<div class="fo-metric">
	<div class="fo-mval">92.9</div>
	<div class="fo-mlbl">F1 Weighted</div>
	</div>
	<div class="fo-metric">
	<div class="fo-mval">77.6</div>
	<div class="fo-mlbl">BERTScore F1</div>
	</div>
	</div>

	<details>
	<summary>Parent Model Comparison</summary>
	<div class="fo-drop">
	<p>The merge trades a small amount of Multimodal accuracy for a <strong>massive improvement</strong> in text-only performance — turning a Multimodal-only model into a genuinely dual-mode classifier.</p>
	<table class="fo-cmp">
	<thead>
	<tr>
	<th>Model</th>
	<th>MM Acc</th>
	<th>MM F1w</th>
	<th>MM BERT</th>
	<th>MM ROUGE</th>
	<th>Text Acc</th>
	<th>Text F1w</th>
	<th>Text BERT</th>
	<th>Text ROUGE</th>
	</tr>
	</thead>
	<tbody>
	<tr>
	<td>Multimodal</td>
	<td class="fo-cmp-warm">92.5</td>
	<td class="fo-cmp-warm">92.3</td>
	<td class="fo-cmp-warm">80.0</td>
	<td class="fo-cmp-warm">38.7</td>
	<td class="fo-cmp-dim">77.5</td>
	<td class="fo-cmp-dim">70.8</td>
	<td class="fo-cmp-dim">73.1</td>
	<td class="fo-cmp-dim">19.0</td>
	</tr>
	<tr>
	<td>Text Adapter</td>
	<td class="fo-cmp-dim">1.5</td>
	<td class="fo-cmp-dim">2.8</td>
	<td class="fo-cmp-dim">69.6</td>
	<td class="fo-cmp-dim">16.2</td>
	<td class="fo-cmp-warm">84.0</td>
	<td class="fo-cmp-warm">89.0</td>
	<td class="fo-cmp-warm">80.7</td>
	<td class="fo-cmp-warm">41.6</td>
	</tr>
	<tr class="fo-cmp-row-merge">
	<td><strong>SLERP Merge</strong></td>
	<td class="fo-cmp-sage">88.5</td>
	<td class="fo-cmp-sage">89.4</td>
	<td class="fo-cmp-sage">77.3</td>
	<td class="fo-cmp-sage">31.6</td>
	<td class="fo-cmp-sage">91.0</td>
	<td class="fo-cmp-sage">92.9</td>
	<td class="fo-cmp-sage">77.6</td>
	<td class="fo-cmp-sage">31.5</td>
	</tr>
	</tbody>
	</table>
	<div class="fo-note fo-note--sage">
	<strong>✶ Key takeaway:</strong> The merge <em>surpasses</em> the text adapter's own accuracy (91.0 vs 84.0) while retaining 95.7% of the Multimodal model's classification accuracy.
	</div>
	</div>
	</details>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 05 · Merge Details ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">05</span>
	<span class="fo-stitle">Merge Details</span>
	</div>
	<div class="fo-sbody">
	<h3 class="fo-sub">Method</h3>
	<div class="fo-card" style="margin-bottom:14px;">
	<div class="fo-row"><span class="fo-k">Method</span><span class="fo-v">SLERP (Spherical Linear Interpolation)</span></div>
	<div class="fo-row"><span class="fo-k">Weight</span><span class="fo-v">0.30 multimodal / 0.70 text</span></div>
	<div class="fo-row"><span class="fo-k">Selection</span><span class="fo-v">Best blend from WS3 sweep</span></div>
	<div class="fo-row"><span class="fo-k">Output</span><span class="fo-v">Full merged weights (no adapter)</span></div>
	</div>
	<h3 class="fo-sub">Process</h3>
	<div class="fo-card fo-card--muted" style="margin-bottom:14px;">
	<div class="fo-row"><span class="fo-v">Both parent LoRA adapters (Multimodal + text) were first merged into the base model (<code>KomdigiITS-8B-DFK-CPT</code>) to produce two full-weight checkpoints. These were then interpolated via SLERP at various weight ratios, with each blend evaluated on both Multimodal and text-only benchmarks to find the optimal trade-off.</span></div>
	</div>
	<h3 class="fo-sub">Label Classes</h3>
	<div class="fo-grid">
	<div class="fo-card">
	<div class="fo-chead">Multimodal Mode · 4 Classes</div>
	<div class="fo-row"><span class="fo-k">Netral</span><span class="fo-v">Factual / non-DFK — no violation</span></div>
	<div class="fo-row"><span class="fo-k">Disinfo</span><span class="fo-v">Claims contradicting established facts</span></div>
	<div class="fo-row"><span class="fo-k">Fitnah</span><span class="fo-v">False claims targeting an individual</span></div>
	<div class="fo-row"><span class="fo-k">Ujrn Kbnci</span><span class="fo-v">Hate speech targeting SARA identity</span></div>
	</div>
	<div class="fo-card fo-card--sage">
	<div class="fo-chead">Text Mode · 5 Classes</div>
	<div class="fo-row"><span class="fo-k">Non-DFK</span><span class="fo-v">Content unrelated to DFK categories</span></div>
	<div class="fo-row"><span class="fo-k">Fakta</span><span class="fo-v">Factual content, verified true</span></div>
	<div class="fo-row"><span class="fo-k">Disinfo</span><span class="fo-v">Claims contradicting established facts</span></div>
	<div class="fo-row"><span class="fo-k">Fitnah</span><span class="fo-v">False claims targeting an individual</span></div>
	<div class="fo-row"><span class="fo-k">Ujrn Kbnci</span><span class="fo-v">Hate speech targeting SARA identity</span></div>
	</div>
	</div>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 06 · Input Formats ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">06</span>
	<span class="fo-stitle">Input Formats</span>
	</div>
	<div class="fo-sbody">
	<p>Both modes use the <code>ministral_3</code> chat template (<code>[INST]</code> / <code>[/INST]</code> delimiters). The default Ministral system prompt is included when no explicit system message is provided.</p>

	<details>
	<summary>Multimodal Mode · dfk_vlm_dataset_v3</summary>
	<div class="fo-drop">
	<p>Image-based classification with contextual metadata. Note: the instruction and context fields are <strong>directly concatenated</strong> with no separator.</p>
	<pre><code><s>[SYSTEM_PROMPT]...default Ministral system prompt...[/SYSTEM_PROMPT][INST]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}[IMG][/INST]Label: {label}

	Analisis: {analisis}</s></code></pre>

	<h3 class="fo-sub">Input Fields</h3>
	<div class="fo-card" style="margin-bottom:14px;">
	<div class="fo-row"><span class="fo-k">Ringkasan</span><span class="fo-v">Content summary. Concatenation of image caption and user-provided text.</span></div>
	<div class="fo-row"><span class="fo-k">Klaim</span><span class="fo-v">Core claim extracted from the content, used as a web search query for fact-checking.</span></div>
	<div class="fo-row"><span class="fo-k">Fakta</span><span class="fo-v">Verification context from web search. Defaults to <code>"Tidak ditemukan sumber yang valid."</code> if none found.</span></div>
	<div class="fo-row"><span class="fo-k">[IMG]</span><span class="fo-v">Screenshot of the social media post being analyzed.</span></div>
	</div>
	<h3 class="fo-sub">Output Fields</h3>
	<div class="fo-card">
	<div class="fo-row"><span class="fo-k">Label</span><span class="fo-v">One of <code>netral</code>, <code>disinformasi</code>, <code>fitnah</code>, or <code>ujaran kebencian</code>.</span></div>
	<div class="fo-row"><span class="fo-k">Analisis</span><span class="fo-v">Free-form Indonesian-language reasoning for the classification.</span></div>
	</div>
	</div>
	</details>

	<details class="fo-det--sage">
	<summary>Text-Only Mode · dfk_text_dataset</summary>
	<div class="fo-drop">
	<p>Text-only classification with explicit system prompt and article references. Uses <code>merge_labels=false</code> (5 classes).</p>
	<pre><code><s>[SYSTEM_PROMPT]Anda adalah sistem deteksi konten DFK berbasis artikel rujukan. Tugas Anda adalah membandingkan klaim dengan artikel rujukan, lalu mengklasifikasikan teks ke dalam salah satu label: Fakta, Disinformasi, Fitnah, Ujaran Kebencian, atau Non-DFK. Jawab dengan format: Label: NamaLabel. penjelasan: ...[/SYSTEM_PROMPT][INST]{klaim}

	Artikel Rujukan: {fakta}[/INST]Label: {label}. penjelasan: {analisis}</s></code></pre>

	<h3 class="fo-sub">Input Fields</h3>
	<div class="fo-card" style="margin-bottom:14px;">
	<div class="fo-row"><span class="fo-k">Klaim</span><span class="fo-v">The claim text to be verified, from the <code>input</code> column (before <code>Artikel Rujukan:</code>).</span></div>
	<div class="fo-row"><span class="fo-k">Fakta</span><span class="fo-v">Reference articles for fact-checking, from the <code>input</code> column (after <code>Artikel Rujukan:</code>).</span></div>
	</div>
	<h3 class="fo-sub">Output Fields</h3>
	<div class="fo-card">
	<div class="fo-row"><span class="fo-k">Label</span><span class="fo-v">One of <code>Fakta</code>, <code>Disinformasi</code>, <code>Fitnah</code>, <code>Ujaran Kebencian</code>, or <code>Non-DFK</code>.</span></div>
	<div class="fo-row"><span class="fo-k">Penjelasan</span><span class="fo-v">Indonesian-language explanation for the classification decision.</span></div>
	</div>
	</div>
	</details>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 07 · Model Sources ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">07</span>
	<span class="fo-stitle">Model Sources</span>
	</div>
	<div class="fo-sbody">
	<div class="fo-card fo-card--muted">
	<div class="fo-row"><span class="fo-k">Framework</span><span class="fo-v"><a href="https://github.com/aitf-its-tim3-dfk/SITA">SITA</a></span></div>
	<div class="fo-row"><span class="fo-k">Base model</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-CPT">aitf-komdigi/KomdigiITS-8B-DFK-CPT</a></span></div>
	<div class="fo-row"><span class="fo-k">Multimodal parent</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-MultimodalClassification">KomdigiITS-8B-DFK-MultimodalClassification</a></span></div>
	<div class="fo-row"><span class="fo-k">Text parent</span><span class="fo-v"><a href="https://huggingface.co/aitf-komdigi/KomdigiITS-8B-DFK-TextClassification">KomdigiITS-8B-DFK-TextClassification</a></span></div>
	</div>
	</div>
	</div>

	<div class="fo-div"></div>

	<!-- ─── 08 · Framework Versions ─── -->
	<div class="fo-section">
	<div class="fo-shead">
	<span class="fo-snum">08</span>
	<span class="fo-stitle">Framework Versions</span>
	</div>
	<div class="fo-sbody">
	<div class="fo-card fo-card--muted">
	<div class="fo-row"><span class="fo-k">Transformers</span><span class="fo-v">5.5.0</span></div>
	<div class="fo-row"><span class="fo-k">PyTorch</span><span class="fo-v">2.11.0+cu128</span></div>
	<div class="fo-row"><span class="fo-k">Unsloth</span><span class="fo-v">2026.5.5</span></div>
	</div>
	</div>
	</div>

	</div>
	</body>
	</html>