README.md

dfk-text-classifier

A Modal deployment of aitf-komdigi/KomdigiITS-8B-DFK-TextClassification — a fine-tuned 8.9B Mistral3 model served as a GPU-backed FastAPI endpoint for Indonesian social media content analysis.

The model classifies content into DFK categories (Disinformasi, Fitnah, Kebencian) and can also summarize text in Bahasa Indonesia.

Features

• DFK classification — detects disinformation, slander, and hate speech from structured social media post metadata
• Summarization mode — summarizes Indonesian text using the base model's general capability (no classification bias)
• Multi-trial MTLA voting — runs N generation trials, scores each via logit-based confidence (K=10 tokens), then majority-votes the result
• Greedy mode — temperature: 0 with single trial for fastest deterministic inference
• CPU memory snapshot — model weights snapshotted after first load for faster cold start on subsequent requests
• JSON sanitizer — middleware that fixes copy-pasted text containing literal newline characters inside JSON strings

Setup

pip install modal
modal setup

If the HF repo is private, create a Modal secret:

modal secret create huggingface-secret HF_TOKEN=hf_your_token_here

Then add secrets=[modal.Secret.from_name("huggingface-secret")] to @app.cls(...) in modal_dfk_v3.py.

Commands

# Dev server (hot-reload, temporary endpoint URL printed to console)
modal serve modal_dfk_v3.py

# Production deploy
modal deploy modal_dfk_v3.py

# Stream logs
modal app logs dfk-text-classification-v3

API

Endpoint: POST https://gghafar--dfk-text-classification-v3-dfkmodel-serve.modal.run

DFK Classification

{
  "ringkasan": "Summary of the social media post",
  "klaim": "Claim made in the post",
  "fakta": "Verified fact for comparison",
  "image_url": "https://...",
  "max_new_tokens": 128,
  "temperature": 0.0,
  "num_trials": 3
}

Response:

{
  "label": "DISINFORMASI",
  "label_key": "disinformasi",
  "description": "Informasi yang menyesatkan",
  "confidence": 97.7,
  "consistency": "3/3",
  "ambiguous": false,
  "reasoning": "...",
  "method": "Unsloth LogitsScore K=10, N=3",
  "trials": [
    { "trial": 1, "label": "DISINFORMASI", "confidence": 97.7, "reasoning": "..." },
    { "trial": 2, "label": "DISINFORMASI", "confidence": 97.7, "reasoning": "..." },
    { "trial": 3, "label": "DISINFORMASI", "confidence": 97.7, "reasoning": "..." }
  ]
}

Summarization

{
  "text": "Teks yang ingin diringkas...",
  "temperature": 0.3
}

Response:

{
  "summary": "Ringkasan teks...",
  "original_length": 668,
  "summary_length": 312
}

Classification Input Fields

Field	Type	Required	Default	Description
ringkasan	string	yes	—	Summary/context of the social media post
klaim	string	yes	—	The specific claim or statement to classify
fakta	string	yes	—	Verified fact(s) to compare the claim against
image_url	string	no	null	Optional image URL for additional context
max_new_tokens	int	no	128	Max tokens to generate (32–1024)
temperature	float	no	0.0	Sampling temperature. 0 = greedy. If > 0 and num_trials > 1, enables MTLA voting
num_trials	int	no	3	Number of generation trials for voting (1–10). Auto-raises temperature to 0.3 when temperature: 0 and num_trials > 1

DFK Labels

Label	Description
Fakta	Content consistent with verified facts
Disinformasi	Misleading or inaccurate information
Fitnah	Serious accusations without verifiable evidence
Ujaran Kebencian	Content attacking or degrading individuals/groups
Non-DFK	Content outside DFK categories

Architecture

File: modal_dfk_v3.py

Component	Description
DFKModel	Modal class on L4 GPU. Loads model via @modal.enter(snap=True), builds FastAPI routes via @modal.enter() (always fresh, never stale from snapshot).
POST /classify	Structured DFK classification with MTLA multi-trial voting.
POST /summarize	Indonesian text summarization using a summarization system prompt.
_mtla_confidence	Computes logit-based confidence score from first K generated token probabilities.
_parse_output	Extracts [LABEL], [CONFIDENCE], and [REASONING] blocks from raw model output.
HF Volume cache	modal.Volume named dfk-8b-cache persists downloaded weights across cold starts.

Model loading flow:

1. snap=True — downloads weights (cached in Volume), loads model + tokenizer via Unsloth FastLanguageModel.from_pretrained in bfloat16 → snapshot taken
2. snap=False (default @modal.enter) — rebuilds FastAPI app with fresh endpoint handlers so code changes take effect immediately after every redeploy

Generation Parameters

Parameter	Default	Description
max_new_tokens	128	Maximum tokens to generate
temperature	0.0	0 = greedy (single trial). > 0 = sampling with MTLA voting
num_trials	3	Number of parallel generation trials for majority voting
repetition_penalty	1.05	Penalizes repeated tokens (hardcoded)

Infrastructure

Setting	Value
GPU	NVIDIA L4 (24 GB VRAM)
CPU	4 vCPU
Memory	24 GB RAM
Timeout	600s
Scale-down	300s idle
Precision	bfloat16 (full size, no quantization)
Snapshot	CPU memory snapshot enabled
Model	aitf-komdigi/KomdigiITS-8B-DFK-TextClassification
Parameters	8.9B