IndoBERTindak

IndoBERTindak

IndoBERTindak is a fine-tuned IndoBERT model for detecting observable behavioral and linguistic signals in Indonesian digital discourse.

The model performs multilabel text classification, meaning that a single text may contain more than one signal, or none of the five signals.

It is built on indobenchmark/indobert-base-p1.

Model Overview

IndoBERTindak detects five observable signals:

These labels represent observable textual behavior.

They should not be interpreted as direct measurements of latent psychological characteristics such as trust, motivation, personality, intelligence, competence, or learning ability.

---

Quick Start

Installation

pip install transformers torch

Load the model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

MODEL_ID = "surodipoikromo/IndoBERTindak"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)

model.eval()

Run multilabel classification

text = "Bagaimana cara menggunakan fitur ini?"

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    max_length=256
)

with torch.no_grad():
    logits = model(**inputs).logits

probabilities = torch.sigmoid(logits)[0]

IndoBERTindak uses label-specific decision thresholds tuned on the development set:

thresholds = torch.tensor([
    0.45,  # N1 — EXPLICIT_HELP_SEEKING
    0.70,  # N2 — ACTIONABLE_KNOWLEDGE_PROVISION
    0.35,  # N3 — LEARNER_SELF_REPORT
    0.35,  # N4 — POSITIVE_CONTENT_FEEDBACK
    0.85,  # N5 — OTHER_DIRECTED_ENCOURAGEMENT
])

predictions = probabilities >= thresholds

Display the results:

labels = [
    "EXPLICIT_HELP_SEEKING",
    "ACTIONABLE_KNOWLEDGE_PROVISION",
    "LEARNER_SELF_REPORT",
    "POSITIVE_CONTENT_FEEDBACK",
    "OTHER_DIRECTED_ENCOURAGEMENT",
]

for label, probability, prediction in zip(
    labels,
    probabilities,
    predictions
):
    print(
        f"{label:32s} "
        f"probability={probability.item():.4f} "
        f"predicted={bool(prediction.item())}"
    )

Because this is a multilabel classifier, do not use argmax() to obtain predictions. Each label is evaluated independently using a sigmoid probability and its corresponding decision threshold.

A text may therefore receive:

• one label,
• several labels, or
• no positive label.

---

Complete Inference Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

MODEL_ID = "surodipoikromo/IndoBERTindak"

LABELS = [
    "EXPLICIT_HELP_SEEKING",
    "ACTIONABLE_KNOWLEDGE_PROVISION",
    "LEARNER_SELF_REPORT",
    "POSITIVE_CONTENT_FEEDBACK",
    "OTHER_DIRECTED_ENCOURAGEMENT",
]

THRESHOLDS = torch.tensor([
    0.45,
    0.70,
    0.35,
    0.35,
    0.85,
])

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
model.eval()


def predict(text):
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=256
    )

    with torch.no_grad():
        logits = model(**inputs).logits

    probabilities = torch.sigmoid(logits)[0]
    predictions = probabilities >= THRESHOLDS

    results = []

    for label, probability, prediction in zip(
        LABELS,
        probabilities,
        predictions
    ):
        results.append({
            "label": label,
            "probability": round(probability.item(), 4),
            "predicted": bool(prediction.item())
        })

    return results


text = "Bagaimana cara menggunakan fitur ini?"

for result in predict(text):
    print(result)

---

Input Format

Top-level text

For independent or top-level comments, provide the target text directly:

<target_text>

Example:

Bagaimana cara melakukan instalasinya?

Replies with parent context

When the target is a reply and the parent comment is available, the model was developed with the following contextual format:

[TARGET] <target_text> [PARENT] <parent_text>

Example:

[TARGET] Coba masuk ke pengaturan lalu aktifkan fiturnya. [PARENT] Bagaimana cara mengaktifkan fitur ini?

The target text remains the primary evidence. Parent context should only be used when necessary to resolve the function or referent of a reply.

---

Decision Thresholds

The final decision thresholds were tuned using the development set only and locked before evaluation on the held-out test set.

Using a single universal threshold such as 0.50 is therefore not recommended when reproducing the reported model performance.

---

Training Data

The model was developed using the Development Gold600 dataset consisting of authentic Indonesian digital-discourse traces.

Dataset composition:

Additional characteristics:

• 600 unique source IDs
• five observable multilabel signals
• Indonesian-language digital discourse
• held-out test data were not used for threshold tuning

The development corpus combines direct assistant adjudication and codebook-guided/rule-assisted annotation.

It should therefore not be described as an independent two-human-coder gold standard.

---

Fine-Tuning Configuration

---

Held-Out Test Performance

Overall performance on the held-out test set:

Per-Label Performance

Performance should be interpreted together with the relatively small per-label support in the held-out test set.

---

Comparison with Classical Baseline

The strongest classical baseline was TF-IDF + Linear SVM.

Absolute Macro-F1 improvement:

+0.086

IndoBERTindak also achieved higher Micro-F1, Weighted-F1, Exact Match, and Sample Jaccard than the classical baseline.

---

Intended Uses

IndoBERTindak is intended primarily for:

• computational analysis of Indonesian digital discourse;
• identification of observable behavioral or linguistic signals;
• large-scale discourse annotation;
• exploratory social computing research;
• communication and interaction analysis;
• information behavior research;
• aggregate analysis of patterns across collections of Indonesian-language texts.

Predictions are best treated as computational annotations rather than definitive judgments about individual users.

---

Out-of-Scope Uses

The model is not intended for:

• psychological diagnosis;
• personality assessment;
• intelligence or competence assessment;
• individual ranking;
• employee or student evaluation;
• automated disciplinary decisions;
• clinical assessment;
• high-stakes profiling;
• inferring hidden mental states from text.

A detected signal indicates that a particular textual pattern is observable in the input. It does not establish why the author produced that text or what psychological state the author had.

---

Limitations

Several limitations should be considered when using the model.

1. Relatively small development dataset. The model was fine-tuned on 600 annotated traces, with 100 observations reserved for held-out evaluation.

2. Uneven label support. Some signals are less frequent than others. In particular, N2 and N5 have modest support in the held-out test set.

3. Context sensitivity. ACTIONABLE_KNOWLEDGE_PROVISION may depend strongly on conversational function and reply context.

4. Precision-recall trade-offs. OTHER_DIRECTED_ENCOURAGEMENT achieved high precision but relatively low recall under its locked threshold.

5. Domain specificity. The model was developed primarily from Indonesian tutorial/discussion discourse on YouTube. Performance on substantially different domains, platforms, genres, dialects, or highly formal text has not been established.

6. Annotation procedure. Annotation expansion included AI-assisted and rule-guided procedures and should not be represented as an independent two-human-coder gold standard.

7. Observable signals only. Model predictions should not be interpreted as direct evidence of latent psychological characteristics.

Users applying IndoBERTindak to a new domain are encouraged to perform domain-specific validation before drawing substantive conclusions.

---

Ethical Considerations

IndoBERTindak was designed to detect observable textual signals, not to profile individuals.

Recommended practice is to:

• analyze predictions at an aggregate level where possible;
• retain human interpretation for substantive conclusions;
• document the domain and sampling procedure;
• report uncertainty and model limitations;
• validate the model when transferring it to substantially different datasets.

The model should not be used as the sole basis for decisions that materially affect an individual.

---

Reproducibility Notes

To reproduce the reported evaluation behavior:

1. use surodipoikromo/IndoBERTindak;
2. tokenize with a maximum sequence length of 256;
3. apply sigmoid independently to each output logit;
4. use the five label-specific locked thresholds;
5. do not use softmax() or argmax() for final multilabel decisions.

The reported held-out metrics correspond to these locked decision rules.

---

Base Model

IndoBERTindak is fine-tuned from:

indobenchmark/indobert-base-p1

The original IndoBERT model should also be consulted for details about its pretraining corpus, architecture, and underlying limitations.

---

Citation

If you use IndoBERTindak in research, please cite the associated publication when it becomes available.

Until then, the model repository can be referenced as:

@misc{indobertindak2026,
  title        = {IndoBERTindak},
  author       = {surodipoikromo},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/surodipoikromo/IndoBERTindak}}
}

Model repository:

https://huggingface.co/surodipoikromo/IndoBERTindak

---

License

This model is released under the Apache License 2.0.

See the repository license information and the license terms of the underlying IndoBERT base model before redistribution or deployment.

---

Disclaimer

IndoBERTindak provides probabilistic computational predictions. Its outputs should be interpreted in relation to the model's training domain, annotation framework, decision thresholds, and documented limitations.

The presence or absence of a predicted signal should not be treated as definitive evidence about an individual's intentions, abilities, personality, motivation, or psychological state.