DEVELOPER GUIDE
GoiBERT for Vietnamese text classification: what to evaluate before fine-tuning
7 November 2026
GoiBERT accepts raw Vietnamese text without a separate word-segmentation step. The released models are Stage 1 encoder backbones with no task-specific prediction head. To classify your own text, you need labelled examples, a classification head and task-specific training. Release details.
01 / TASK FIT
Start with a clear decision
Routing support tickets or analysing customer reviews are possible projects, not demonstrated deployments. Define the labels and what happens after each prediction. A review such as “Món ngon nhưng giao chậm” contains both praise and criticism: decide whether you need one overall sentiment, multiple topics or sentiment by aspect.
02 / MODEL SELECTION
Choose a model to test
GoiBERT-S (Small) has 26.4M parameters; GoiBERT-M (Medium) has 49.4M. Test S first if memory is constrained, then M under the same conditions. Select on task quality and measured resource use, not parameter count alone.
Both model cards recommend a 256-token input length. Measure how often your inputs exceed that length and inspect what truncation removes. Raw-text support does not establish reliability on dialects, missing accents or unfamiliar domains.
03 / EVALUATION
Set a fair evaluation
Keep related conversations and duplicate reviews in the same data split. Reserve a test set before tuning. Compare a simple baseline, such as TF-IDF with logistic regression, and a relevant Vietnamese encoder using the same labels, splits and evaluation criteria. Record tuning budgets and any extra teacher supervision.
Set your minimum requirements before training: macro-F1, precision and recall for important classes, and maximum acceptable latency and memory on your hardware. Include tokenization and the classification head when timing the full pipeline. Tune on validation data; evaluate the selected setup on the reserved test set.
04 / FAILURE CASES
Inspect failures before use
Review negation, sarcasm, mixed Vietnamese/English, rare labels, conflicting sentiments and long inputs. Decide which uncertain or unfamiliar cases go to a person. Use the model only if it meets those requirements; otherwise revise the labels, data or method.
Before use, read the model’s usage license and linked upstream terms. Start with the model cards and research overview to record the release you evaluate.