A SLERP-merged full model combining two specialized LoRA adapters into a single unified checkpoint. The merge blends the 8B Multimodal Multimodal classifier with the text-only DFK classifier, both built on KomdigiITS-8B-DFK-CPT (Ministral-3-8B-Base-2512).

The result is a model that operates in two modes: image-based multimodal classification and text-only classification — strong performance on both tasks from a single set of weights.

✶ Note: Best SLERP blend from a parameter sweep (w=0.30 multimodal, w=0.70 text). Parent adapters were merged into base model weights, producing a standalone model with no adapter overhead.

02 Model Details

Identity

DevelopedDFK Tim 1 & Tim 3 ITS

TypeMerged model (SLERP)

LanguageIndonesian

ModesMultimodal + Text-only

Architecture

BaseKomdigiITS-8B-DFK-CPT

ArchMistral3ForConditionalGeneration

Params8B (full merged weights)

Precisionfloat16

Parent Models

Multimodal Parent · w=0.30

Model8B-DFK-MultimodalClassification

TaskMultimodal — image + text

Datasetdfk_vlm_dataset_v3

Text Parent · w=0.70

Model8B-DFK-TextClassification

TaskText-only DFK classification

Datasetdfk_text_dataset

03 Uses

Direct Use

This model supports two input modes from a single checkpoint:

Mode 1 · Multimodal

Image-based content moderation. Given a social media screenshot with contextual metadata, classifies into 4 labels: netral, disinformasi, fitnah, or ujaran kebencian.

Mode 2 · Text-Only

Text-based DFK detection using article references. Classifies into 5 labels: Fakta, Disinformasi, Fitnah, Ujaran Kebencian, or Non-DFK.

Out-of-Scope Use

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.

04 Evaluation

Evaluated on held-out validation splits using greedy decoding (temperature=0.0) and BERTScore (bert-base-multilingual-cased).

Multimodal Task (Image + Text)

88.5

Accuracy

89.4

F1 Weighted

77.3

BERTScore F1

Text-Only Task

91.0

Accuracy

92.9

F1 Weighted

77.6

BERTScore F1

Parent Model Comparison

The merge trades a small amount of Multimodal accuracy for a massive improvement in text-only performance — turning a Multimodal-only model into a genuinely dual-mode classifier.

Model	MM Acc	MM F1w	MM BERT	MM ROUGE	Text Acc	Text F1w	Text BERT	Text ROUGE
Multimodal	92.5	92.3	80.0	38.7	77.5	70.8	73.1	19.0
Text Adapter	1.5	2.8	69.6	16.2	84.0	89.0	80.7	41.6
SLERP Merge	88.5	89.4	77.3	31.6	91.0	92.9	77.6	31.5

✶ Key takeaway: The merge surpasses the text adapter's own accuracy (91.0 vs 84.0) while retaining 95.7% of the Multimodal model's classification accuracy.

05 Merge Details

Method

MethodSLERP (Spherical Linear Interpolation)

Weight0.30 multimodal / 0.70 text

SelectionBest blend from WS3 sweep

OutputFull merged weights (no adapter)

Process

Both parent LoRA adapters (Multimodal + text) were first merged into the base model (KomdigiITS-8B-DFK-CPT) 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.

Label Classes

Multimodal Mode · 4 Classes

NetralFactual / non-DFK — no violation

DisinfoClaims contradicting established facts

FitnahFalse claims targeting an individual

Ujrn KbnciHate speech targeting SARA identity

Text Mode · 5 Classes

Non-DFKContent unrelated to DFK categories

FaktaFactual content, verified true

DisinfoClaims contradicting established facts

FitnahFalse claims targeting an individual

Ujrn KbnciHate speech targeting SARA identity

06 Input Formats

Both modes use the ministral_3 chat template ([INST] / [/INST] delimiters). The default Ministral system prompt is included when no explicit system message is provided.

Multimodal Mode · dfk_vlm_dataset_v3

Image-based classification with contextual metadata. Note: the instruction and context fields are directly concatenated with no separator.

<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>

Input Fields

RingkasanContent summary. Concatenation of image caption and user-provided text.

KlaimCore claim extracted from the content, used as a web search query for fact-checking.

FaktaVerification context from web search. Defaults to "Tidak ditemukan sumber yang valid." if none found.

[IMG]Screenshot of the social media post being analyzed.

Output Fields

LabelOne of netral, disinformasi, fitnah, or ujaran kebencian.

AnalisisFree-form Indonesian-language reasoning for the classification.

Text-Only Mode · dfk_text_dataset

Text-only classification with explicit system prompt and article references. Uses merge_labels=false (5 classes).

<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: [object Object] penjelasan: {analisis}</s>

Input Fields

KlaimThe claim text to be verified, from the input column (before Artikel Rujukan:).

FaktaReference articles for fact-checking, from the input column (after Artikel Rujukan:).

Output Fields

LabelOne of Fakta, Disinformasi, Fitnah, Ujaran Kebencian, or Non-DFK.

PenjelasanIndonesian-language explanation for the classification decision.

07 Model Sources

FrameworkSITA

Base modelaitf-komdigi/KomdigiITS-8B-DFK-CPT

Multimodal parentKomdigiITS-8B-DFK-MultimodalClassification

Text parentKomdigiITS-8B-DFK-TextClassification

08 Framework Versions

Transformers5.5.0

PyTorch2.11.0+cu128

Unsloth2026.5.5

aitf-its-tim3-dfk/ministral-8b-merged-ws3

KomdigiITS-8B-DFKMerged Classification

Parent Models

Direct Use

Out-of-Scope Use

Multimodal Task (Image + Text)

Text-Only Task

Method

Process

Label Classes

Input Fields

Output Fields

Input Fields

Output Fields

aitf-its-tim3-dfk/
ministral-8b-merged-ws3

KomdigiITS-8B-DFK
Merged Classification