A continual-learning MLOps system for Indonesian sentiment analysis. When the language in new data changes (drift), the system retrains the model with LoRA + replay, so it learns new slang without forgetting old data.
Done: data pipeline (DVC), baselines, the LoRA + replay method, a prediction API with Docker and CI, the retraining pipeline (drift check, Kaggle GPU, promotion rule, registry, API switch) run end to end on real services, monitoring with Prometheus and Grafana, a public demo, and a separate abusive language flag in the API.
Try it: https://huggingface.co/spaces/amc76/slangguard-demo. The model runs in your browser, so nothing you type is sent anywhere. See docs/design.md for the full design and roadmap.
Modeling
Data and experiments
Drift and pipeline
Serving
Quality and tooling
Monitoring
Demo
Badges are made with shields.io in the style of ileriayo/markdown-badges.
Language change is simulated: four periods, from formal reviews to tweets that are increasingly rewritten with slang. P3 and P4 are not real future language. Mean over 3 seeds, macro-F1 on the test sets of all four periods after the last update:
| Method | AvgF1 | P1 (oldest period) | Training time of the last update |
|---|---|---|---|
| Naive fine-tune (new data only) | 0.677 | 0.759 | 32 s |
| Full retrain on all data | 0.724 | 0.925 | 328 s |
| LoRA + 5% replay | 0.725 | 0.900 | 29 s |
Details and limits: docs/results.md and docs/baselines.md. The trained adapter is on Hugging Face: amc76/slangguard-indobert-lora-replay.
You need Docker. The first start downloads the model (about 480 MB, a few minutes), and later starts take seconds.
docker compose upThen, in another terminal:
curl -X POST localhost:8000/predict -H 'Content-Type: application/json' \
-d '{"texts": ["makanannya enak banget, pelayanannya ramah", "parah sih, lama banget nunggunya"]}'The answer has a sentiment label and the three class probabilities for each text, in the same
order, and a separate flag for abusive language (abusive and abusive_score), see below.
Interactive documentation is at http://localhost:8000/docs, and GET /health shows the model
version. Stop with docker compose down. Without Docker:
uv run uvicorn src.serving.app:create_default_app --factory.
One sentence takes about 60 ms in Docker on a laptop CPU, and the container uses about 1 GB.
Sentiment is not rudeness: "kontol" alone can score as neutral in sentiment. So a second small
model (TF-IDF over letter groups and logistic regression, 1.3 MB) flags abusive words, and the
API returns it next to the sentiment: "abusive": true, "abusive_score": 1.0. F1 is 0.893 on its
own test set. Its first version flagged plain praise like "enak banget", which only showed up when
it was checked on this project's own text, and the fix is described in
docs/abusive.md. The model is CC BY-NC-SA 4.0 because its data is:
amc76/slangguard-abusive. The demo in the
browser does not show it yet.
The API reports request counts, latency, the mix of predicted labels and word pieces per word (a live sign that slang is arriving). It is optional and does not start with the API:
docker compose --profile monitoring up # Grafana at http://localhost:3000, Prometheus at :9090
Details, the metrics and what was measured: docs/monitoring.md.
The page at https://huggingface.co/spaces/amc76/slangguard-demo loads an 8-bit version of the model
into your browser (about 125 MB on the first visit). It is a static page because Gradio Spaces now
need a paid plan. How it was built and checked against the Python model: docs/demo.md.
To try the model with Gradio on your own computer: PYTHONPATH=. uv run python demo/app.py.
New text is checked for drift with Evidently. If the language changed, a Prefect flow trains a copy of the current model on a Kaggle GPU, compares it with the current one on validation sets, and switches the API only if it passes. How it works, how it was calibrated, and what has been run for real: docs/pipeline.md.
uv sync # all dependencies
uv run ruff check . && uv run ruff format --check . && uv run pytestTraining runs on Kaggle (see notebooks/), and the data is tracked with DVC.