.svg){kind=small}
Understanding Training Settings
When fine-tuning AI models, a few key settings determine how well the model learns and how long training takes. Think of these settings as dials you can adjust to balance quality, speed, and cost.
Good News for You
Bios provides intelligent defaults for all training settings, optimized through extensive research. You can start training immediately with settings that work well for most use cases—no expertise required.
Learning Speed: How Fast the Model Adapts
The learning rate controls how quickly the model changes as it learns from your examples. It's one of the most important settings affecting training success.
Too Slow
The model learns very gradually. Training takes longer and may not reach optimal performance within your timeframe.
Just Right
The model learns efficiently, making steady progress without instability. This is what Bios automatically provides.
Too Fast
The model changes too rapidly, potentially becoming unstable or "forgetting" what it learned. Training may fail.
Automatic Optimization
Bios automatically calculates the optimal learning speed for each model based on extensive research and testing. The platform has been validated across hundreds of training scenarios to ensure you start with settings that work.
How Bios Determines the Right Speed
The optimal learning speed depends on several factors that Bios considers automatically:
- →Model Size: Larger models typically need different learning speeds than smaller ones
- →Training Method: Whether you're using efficient fine-tuning or full model updates
- →Model Architecture: Different expert models have been tested to find their ideal settings
Batch Size: Balancing Quality and Speed
Batch size determines how many examples the model learns from at once before updating. This setting creates a trade-off between training quality and training speed.
The Trade-off Explained
Smaller Batches (32-64 examples)
✓Often produces the highest quality results
✓Model learns more nuanced patterns
⚠Takes longer to complete training
Larger Batches (256-512 examples)
✓Training completes much faster
✓More efficient use of resources
⚠May sacrifice some final quality
Recommended: Batch Size of 128
Based on extensive testing, a batch size of 128 provides an excellent balance between quality and speed for most use cases. This is Bios's default setting.
When to consider adjusting:
- • Use smaller batches (64) if achieving the absolute best quality is your priority
- • Use larger batches (256) if you need results quickly and can accept slightly lower quality
Training Duration
Regardless of batch size, ensure your training runs long enough for the model to learn effectively:
Minimum Steps
Sufficient for simple adaptations
Recommended Steps
For best results on complex tasks
Model Capacity: How Much the Model Can Learn
When using efficient fine-tuning (the default approach), the "rank" setting determines how much new information the model can absorb. Think of it as the model's learning capacity.
Lower Capacity
Best for: Small datasets, simple adaptations, or when using preference-based training methods.
Recommended
Best for: Most supervised learning tasks. Provides good capacity without excessive resource use.
Higher Capacity
Best for: Large datasets, complex tasks, or when the model needs to learn extensive new knowledge.
Capacity and Dataset Size
As a general rule, your model's capacity should match your dataset size. Larger datasets benefit from higher capacity settings, while smaller datasets work well with lower capacity. Bios defaults to 32, which works well for most enterprise use cases.
When to Adjust These Settings
Bios's default settings work well for most scenarios. However, certain situations may benefit from adjustments.
Signs Your Training Needs Adjustment
Training Becomes Unstable
If the model's performance fluctuates wildly or training fails, try reducing the learning speed.
Progress Stops Too Early
If the model stops improving before reaching good performance, you may need to increase learning speed or capacity.
Training Takes Too Long
If you need faster results, increase the batch size (though this may slightly reduce quality).
Quality Isn't High Enough
For maximum quality, decrease batch size and increase capacity, though training will take longer.
Testing Different Settings
If you want to experiment with settings to find what works best for your specific use case, try a small number of variations:
- 1.Start with Bios's defaults and run a training session
- 2.If results aren't satisfactory, try adjusting one setting at a time
- 3.Compare results to understand which changes help your specific case
Quick Reference Guide
| Setting | Default | What It Does | When to Change |
|---|---|---|---|
| Learning Speed | Auto-optimized | Controls how quickly the model learns | Training unstable or stuck |
| Batch Size | 128 | Balances training speed and quality | Need faster/higher quality |
| Model Capacity | 32 | How much new info model can learn | Very large/small datasets |
| Training Steps | 1000+ | How long training runs | Based on dataset size |
The Bottom Line
Bios handles the complexity of training configuration for you. Start with the defaults—they're optimized for success based on extensive research and testing across diverse use cases.
Most users achieve excellent results without ever adjusting these settings. If you do need to fine-tune, focus on one setting at a time and measure the impact before making additional changes.