Repository navigation
Training Sequence
Sudhanshu edited this page Nov 11, 2026
·
1 revision
The full fine-tuning path is implemented before the frozen-backbone comparison.
sequenceDiagram
autonumber
participant User as Python/Jupyter
participant Hub as Hugging Face Hub
participant DS as Dataset + Collator
participant Model as DistilBERT Classifier
participant Trainer as Trainer
participant Disk as Local Files
User->>DS: Build 30 records and stratified split
User->>Hub: Load tokenizer and pretrained weights
Hub-->>User: Tokenizer and weights
User->>DS: Tokenize and configure dynamic padding
User->>Model: Initialize four-label classifier
User->>Trainer: TrainingArguments + datasets + model
loop For 4 epochs
Trainer->>DS: Read padded batches
DS-->>Trainer: input_ids, attention_mask, labels
Trainer->>Model: Forward / loss / backward / optimizer step
Trainer->>Model: Evaluate on 6 validation samples
Model-->>Trainer: logits / metrics
Trainer->>Disk: Save eligible epoch checkpoint
end
Trainer->>Disk: Save best_model and tokenizer
User->>Trainer: predict(validation), evaluate()
Trainer-->>User: accuracy, weighted F1, report and matrix
The Trainer calculates training loss and optimization updates. Evaluation runs each epoch; saving is also configured per epoch for the full fine-tune. The best model is selected by eval_f1_weighted. The code separately writes the selected model and tokenizer under support_classifier/best_model.
The diagram is an architectural sequence, not a trace from an executed GPU run.
Getting started
Architecture
- System Context
- High-Level Design
- Low-Level Design
- Model Components
- Training Sequence
- Data & Labels
- Tokenization
Training & results
Engineering reference