Repository navigation
Full Fine Tuning
Sudhanshu edited this page Nov 11, 2026
·
1 revision
The full experiment initializes AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, num_labels=4, ...) and trains the entire pretrained encoder plus the classification head.
| Parameter | Source value |
|---|---|
| Model | distilbert-base-uncased |
| Epochs | 4 |
| Per-device train/eval batch size | 8 |
| Learning rate | 2e-5 |
| Warmup fraction | 0.10 |
| Weight decay | 0.01 |
| Evaluation and saving | Every epoch |
| Best-checkpoint metric |
f1_weighted (maximize) |
| Saved checkpoint limit | 2 |
| Full output folder | ./support_classifier |
Depending on hardware, the script selects BF16, FP16, or full-precision execution. CUDA is optional; CPU execution is supported.
The output is a four-class encoder classifier, not a generative foundation model.
Getting started
Architecture
- System Context
- High-Level Design
- Low-Level Design
- Model Components
- Training Sequence
- Data & Labels
- Tokenization
Training & results
Engineering reference