Supervised fine-tuning (SFT) adapts a pretrained model using labeled input–output examples. In full SFT, every model parameter is updated, making task behavior, domain vocabulary, and response patterns part of the model’s defaults. The extra capacity comes with higher memory, training, and monitoring costs than LoRA, which trains small adapters on a frozen base model. section 6.0 section 6.1
Start with the least expensive intervention that can meet the quality target. Prompting or a decode-time schema may be enough for format alone; retrieval-augmented generation (RAG) is better suited to frequently changing knowledge or verifiable citations. Start training with LoRA, then consider full SFT when a well-tuned adapter plateaus on a meaningful quality gap. section 1.6.2 section 1.7 section 5.2.3
Before committing to full SFT, check the decision against production constraints:
- Use a shared held-out evaluation set to compare the base model, prompting, RAG, LoRA, and SFT where applicable. Resolving differences of a few points typically requires 200–500 evaluation examples. section 1.7
- Treat 1,000 curated examples as a floor for testing full SFT, not a guarantee. Plan for it at 10,000 or more; at 100,000 or more, LoRA capacity may become a bottleneck. section 5.2.3 section 6.13
- Budget for GPUs, training time, safety checks, and ongoing maintenance. Full SFT often offers only single-digit percentage improvements over well-tuned LoRA, so measure whether the gain has business value. section 6.4.1 section 6.13
Curate representative, diverse examples with correct, complete assistant responses and consistent conventions. Strip sensitive information and normalize noisy records before converting them to chat-structured JSONL. In the illustrated setup, system and user turns condition the model, while loss is computed only on assistant tokens; inconsistencies in those targets become training signals. section 6.7 section 6.7.1
A practical full-SFT workflow is:
- Prepare domain-filtered training data and separate validation and test examples. Establish a baseline and preserve the test set for final comparisons. section 6.1 section 8.2.7
- Load the pretrained model and train without a LoRA configuration. For 4B–7B models, starting settings include a 2e-5 learning rate, small per-device batches, gradient accumulation, and gradient checkpointing. section 6.1.3 section 6.1.2
- Run a smoke test, monitor training and validation loss and gradient norms, and select a suitable checkpoint. Rising validation loss can indicate overfitting; sudden gradient spikes can signal divergence. section 6.1.3 section 6.8
- Inspect generated answers, compare task metrics on held-out examples, and run safety regression checks. Also test general capabilities: aggressive updates can cause catastrophic forgetting. section 6.1.4 section 6.8
Table 6.1 Starting hyperparameters for SFT on 4B to 7B models
|
Parameter
|
Recommended
|
Notes
|
|
num_train_epochs
|
3
|
Rarely need more than 5; watch val loss
|
|
per_device_train_batch_size
|
1 to 2
|
Limited by GPU VRAM
|
|
gradient_accumulation_steps
|
8 to 16
|
An effective batch of 16 is a good default
|
|
learning_rate
|
2e-5
|
Range: 1e-5 to 5e-5
|
|
warmup_ratio
|
0.03 to 0.10
|
Or 50 to 200 fixed steps
|
|
weight_decay
|
0.01
|
Mild regularization
|
|
max_length
|
512 to 1024
|
Must cover your longest examples
|
|
bf16
|
True
|
Use fp16 only if bf16 unsupported
|
|
gradient_checkpointing
|
True
|
Essential for consumer GPUs
|
|
save_total_limit
|
2 to 3
|
Disk space: intermediate checkpoints
are about 24 GB each with optimizer state; the final model is about 8 GB
|
Starting hyperparameters show the memory and stability choices for full SFT on 4B–7B models. section 6.1.2
IT support assistant
A Qwen3-4B assistant is trained on 450 IT support examples to produce step-by-step answers. On a 50-question held-out test, Token-F1 was 0.153 for the base model, 0.154 for LoRA, and 0.168 for full SFT. The test is small: differences of a few points are within its noise, so this run demonstrates the workflow rather than proving a reliable SFT advantage. The 450-example run is illustrative; the decision floor for testing full SFT remains about 1,000 curated examples. section 6.1.3 section 5.2.3 section 8.2.7 section 6.13
SFT learns from target responses, not explicit comparisons between acceptable answers. If the model can already perform the task but needs subtler preferences for concision, tone, or safety, preference optimization can refine an SFT checkpoint; it is not needed when SFT already meets the quality bar. section 6.10 section 6.10.1


