Model Name: sdd-emotion-kepolisian Base Model: indobenchmark/indobert-base-p2 Task: Emotion classification (6-class) of Indonesian police-related news Language: Indonesian Domain: Police-related news articles

Model Description

Fine-tuned IndoBERT that labels the emotion a news article carries for Polri as an institution — how the article lands when it reaches Polri's media-monitoring desk. This is deliberately not the emotion felt by the public, nor the emotion expressed by the journalist.

That distinction matters and is the most important thing to understand before using this model. A report of public anger directed at the police is not anger here; for the institution it is bad news, so it is sadness (or fear when unrest may follow). Conversely, a story about a solved child abuse case is happy, because for the institution it is good news — even though its subject matter is grim.

Classes:

anger — the institution is attacked: officers assaulted or obstructed, police facilities damaged, provocation, defamation or hoaxes against Polri
fear — security threat or pressure: terror, escalating conflict, unresolved armed crime, potential mass unrest
sadness — bad news or grief for Polri: officers killed, personnel misconduct (arrest, dismissal, ethics tribunal), reputational damage, fatalities in handled incidents
neutral — routine or procedural coverage with no emotional weight for Polri; also any article that does not concern Polri at all
happy — good news: cases solved, successful operations, awards, public appreciation
love — warmth and closeness with the community: social service, officers helping residents, moving moments, tributes

Label provenance: labels were produced by a 3-vote GPT-4o-mini ensemble (rotated few-shot ordering, temperature 0.3), not by human annotators. Only rows where at least 2 of 3 votes agreed were kept. See the optimization report for the manual QA results and the limitations this implies.

Performance Metrics

Evaluated on a frozen test set of 457 articles held fixed across every experiment in this project, so all numbers are directly comparable.

Metric	This checkpoint (seed 42)	Config average (5 seeds)
Accuracy	0.7702	0.7816 +/- 0.0226
Macro F1	0.7080	0.7252 +/- 0.0317
Weighted F1	0.7597	—
Latency p95 (GPU)	7.18 ms	—
Model size	474.7 MB	—

Use the 5-seed average as the expected performance. The single checkpoint figure is one draw from a distribution whose spread is documented above; run-to-run variation on this dataset is real and was measured rather than assumed.

Per-Class Performance (this checkpoint; test set is not class-balanced)

Class	Precision	Recall	F1	Support	Recall across 5 seeds
`anger`	0.5455	0.1875	0.2791	32	0.250 +/- 0.049
`fear`	0.6897	0.6667	0.6780	60	0.643 +/- 0.165
`sadness`	0.7327	0.7400	0.7363	100	0.762 +/- 0.023
`neutral`	0.6800	0.8500	0.7556	100	0.878 +/- 0.036
`happy`	0.8810	0.8810	0.8810	84	0.883 +/- 0.018
`love`	0.9359	0.9012	0.9182	81	0.894 +/- 0.045

anger is the weak class and its weakness is structural, not a tuning failure. The 58k-article corpus contains only ~270 articles that qualify as institutional anger, leaving 191 for training and 32 for testing. At n=32 the recall confidence interval is roughly +/-17 points. Do not rely on this class for automated alerting without human review.

Comparison against the previous production model

	`faizaulia/e5-fine-tune-polri-news-emotion`	This model
Macro F1	0.3284	0.7252
Size	2136 MB	475 MB
Latency p95	11.33 ms	7.18 ms

An important caveat on that comparison: the legacy model was trained on generic Indonesian emotion data, while this test set is labelled with institutional semantics. Part of the gap therefore reflects a difference in what "emotion" means, not only a difference in model quality. Two other generic emotion models scored 0.22 and 0.06 on the same test set, collapsing in the same way — which is the signature of a label-definition mismatch.

Usage

Load Model

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("AzrilFahmiardi/sdd-emotion-kepolisian")
model = AutoModelForSequenceClassification.from_pretrained("AzrilFahmiardi/sdd-emotion-kepolisian")
model.eval()

Inference

python

def predict_emotion(title: str, fulltext: str) -> str:
    text = f"{title} {fulltext}".strip()
    enc = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
    with torch.no_grad():
        logits = model(**enc).logits
    return model.config.id2label[int(logits.argmax(-1))]

# Example
print(predict_emotion(
    "Polres membongkar jaringan narkoba lintas provinsi",
    "Satreskrim menyita 12 kilogram sabu dan menangkap lima tersangka ..."
))
# -> "happy"

Output Format

Returns one of: anger, fear, sadness, neutral, happy, love.

For the production endpoint contract, map to Title Case:

python

PRODUCTION_LABEL_MAP = {
    "anger": "Anger", "fear": "Fear", "sadness": "Sadness",
    "neutral": "Neutral", "happy": "Happy", "love": "Love",
}

Input/Output


Input	`title` + `fulltext`, concatenated with a space, untruncated by the caller
Max length	256 tokens (the tokenizer truncates; long articles are safe)
Output	one of 6 lowercase class names

Preprocessing Note

Pass the whole article — title plus full text — and let the tokenizer truncate at 256 tokens. Do not pre-extract a short excerpt.

max_len 256 was chosen empirically and the result was counter-intuitive: 256 tokens beat 512 by 0.056 Macro F1 even though it truncates about 90% of articles. Indonesian news follows an inverted-pyramid structure, so the emotional framing sits in the lead while the tail carries boilerplate and procedural quotes that dilute the signal. Shorter is better here; 128 tokens was also tested and was worse, so 256 is a genuine optimum rather than a monotonic trend.

This also means the model does not need the legacy 3-sentence extract_and_concatenate preprocessing, and should not be given it.

Limitations

anger is unreliable (recall 0.25 +/- 0.05). Corpus-limited, not fixable by tuning.
Labels are model-generated, not human-annotated. Systematic errors made by the labelling LLM are inherited by this model.
Rare classes were partly retrieved by keyword search, which introduces selection bias in anger, fear, and love. The test set draws from the same pool, so its estimates are not more optimistic than reality but are not a random sample of the corpus either.
Semantics differ from the field name's history. The production field emotion_polisi previously carried generic emotion labels. Class names are unchanged, so the change is invisible to API consumers — anyone comparing time series across the switchover will see a discontinuity.
Trained on articles published January–June 2026 only.