Instructions to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries
PEFT
How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with PEFT:
```
Task type is invalid.
```

How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with Transformers:

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

pipe = pipeline("text-generation", model="aitf-komdigi/KomdigiITS-8B-DFK-CPT")

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("aitf-komdigi/KomdigiITS-8B-DFK-CPT")
model = AutoModelForImageTextToText.from_pretrained("aitf-komdigi/KomdigiITS-8B-DFK-CPT")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aitf-komdigi/KomdigiITS-8B-DFK-CPT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-komdigi/KomdigiITS-8B-DFK-CPT",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/aitf-komdigi/KomdigiITS-8B-DFK-CPT

SGLang

How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT 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-komdigi/KomdigiITS-8B-DFK-CPT" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-komdigi/KomdigiITS-8B-DFK-CPT",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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-komdigi/KomdigiITS-8B-DFK-CPT" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aitf-komdigi/KomdigiITS-8B-DFK-CPT",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Unsloth Studio

How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for aitf-komdigi/KomdigiITS-8B-DFK-CPT to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for aitf-komdigi/KomdigiITS-8B-DFK-CPT to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for aitf-komdigi/KomdigiITS-8B-DFK-CPT to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="aitf-komdigi/KomdigiITS-8B-DFK-CPT",
    max_seq_length=2048,
)

Docker Model Runner
How to use aitf-komdigi/KomdigiITS-8B-DFK-CPT with Docker Model Runner:
```
docker model run hf.co/aitf-komdigi/KomdigiITS-8B-DFK-CPT
```

Ministral-8B-DFK-CPT

Model Details

Model Description

Ministral-8B-DFK-CPT adalah model bahasa Indonesia yang dirancang untuk Continued Pre-Training (CPT) pada domain spesifik DFK (Disinformasi, Fitnah, Kebencian) dan Pengetahuan Umum. Model ini dibangun di atas arsitektur Ministral-3-8B-Base-2512 menggunakan framework Unsloth untuk pelatihan yang lebih cepat dan efisien dengan pendekatan LoRA (Low-Rank Adaptation).

Developed by: aitf-komdigi
Model type: Causal Language Model
Base architecture: Ministral-3-8B-Base-2512
Primary language: Indonesian (id)
License: Apache-2.0

Training Data Composition

Dataset yang digunakan merupakan gabungan dari data spesifik DFK dan data umum (General) dalam bahasa Indonesia dengan target komposisi 80% Domain DFK dan 20% General. Berikut adalah rincian komposisi berdasarkan persentase realita dan jumlah token yang terkumpul:

Komponen	Persentase Realita	Token Terkumpul Ministral 3
Berita (DFK, Formal)	44%	~100.793 juta
Berita (DFK, Nonformal)	19%	~43.421 juta
Berita (DFK, Formal)	6%	~13.798 juta
Wikipedia (General, Formal)	13%	~30.166 juta
Wikipedia (DFK, Formal)	10%	~22.920 juta
Jurnal (General, Formal)	7%	~16.309 juta

Intended Use

Direct Use (Recommended)

Model ini ditujukan untuk Continued Pre-Training, khususnya untuk:

Adaptasi domain identifikasi konten Disinformasi, Fitnah, dan Kebencian di ruang digital Indonesia.
Pengayaan pengetahuan konteks sosiopolitik lokal.
Pre-adaptation sebelum dilanjutkan ke tahap Instruction Tuning atau DPO/RLHF.

Out-of-Scope Use

Chat-oriented instruction following tanpa fine-tuning lanjutan (ini adalah model base/CPT).

How to Get Started

Load the model using HuggingFace Transformers atau Unsloth:

from unsloth import FastLanguageModel
import torch

model_id = "aitf-komdigi/KomdigiITS-8B-DFK-CPT"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    max_seq_length=2048,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)

input_text = "Disinformasi di Indonesia terutama menyebar lewat"
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.7)
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Procedure

Model dilatih menggunakan Continued Pre-Training (CPT) dengan library Unsloth.

Hyperparameters

Precision: bf16 (mixed precision)
Quantization: 4-bit (via Unsloth)
Max Sequence Length: 2048
LoRA Rank (r): 16
LoRA Alpha: 32
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Batch size: 16 / device
Gradient accumulation: 4 (effective batch size = 64)
Learning rate: 5e-5 (cosine schedule)
Warmup steps: 100
Epochs: 1
Optimizer: adamw_8bit

Evaluation

Results

Berdasarkan evaluasi sebelum (baseline) dan sesudah pelatihan:

Baseline Self-PPL (Pure DFK Test Sample): 7.7082
Final Self-PPL (Pure DFK Test Sample): 5.8837
Self Improvement: +23.67% (Model jauh lebih baik mengenali teks berdomain DFK)
Wiki-ID PPL Degradation: -0.30% (Catastrophic forgetting sangat minimal terhadap bahasa umum)

Tech Stack & Hardware

Framework: PyTorch, Unsloth, Hugging Face Transformers, TRL
Monitoring: Weights & Biases (WandB)
Hardware: NVIDIA RTX PRO 6000 Blackwell Server Edition

Downloads last month: 181

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for aitf-komdigi/KomdigiITS-8B-DFK-CPT

Base model

mistralai/Ministral-3-8B-Base-2512

Adapter

(3)

this model

Adapters

4 models

Finetunes

11 models