MLX Training Interface

Training Configuration

Model

Training Dataset

Output Directory

Batch Size

Learning Rate

Max Iterations

Warmup Steps

Max Sequence Length

Weight Decay

LR Decay Factor

Fine-tuning Type Full trains all parameters, LoRA/DoRA use parameter-efficient methods

Layers to Fine-tune Use -1 for all layers or specify number (for LoRA/DoRA)

LoRA/DoRA Parameters: These settings control the adapter training behavior

LoRA Rank Higher rank = more parameters, better quality (8-64 typical)

LoRA Scale Scaling factor for adapter weights (10-30 typical)

LoRA Dropout Dropout rate for regularization (0.0-0.1 typical)

Target Modules Which layers to apply LoRA to

LR Schedule

Save Every X Steps

Eval Every X Steps

Quick Val % Percentage of validation set used for quick checks

Validation Split Fraction of data used for validation

Enable Early Stopping When enabled, training stops if validation loss stops improving (disabled by default)

Training Estimates: Configure parameters to see estimates

Learning Rate Schedule Preview

Training Status

No training in progress

Training Sessions

Loading checkpoints...

Training Session

Select Training Session

Iteration

Train Loss

Val Loss

Perplexity

Tokens/sec

Trained Tokens

Epoch

Memory (GB)

Learning Rate

Warmup Steps

LR Decay

Weight Decay

Training Progress

E.Time: -- R.Time: -- 0%

Top 3 Best Checkpoints

No training data available

All Available Checkpoints

Select Checkpoint to Publish

This list shows all available checkpoints for the selected training session.

Training Sessions

Important: only compare sessions trained on the same dataset.

Loading training sessions...

Training Analysis Dashboard

Training Session Analysis

Select one or more training sessions from the left panel to view their performance metrics and compare results.

Available Analysis:

Loss Curves
Perplexity Evolution
Loss Stability
Generalization Gap

Tips:

View single session metrics
Compare multiple sessions
Use same training dataset
Review hyperparameters

Loss Analysis Training vs Validation

Perplexity Analysis Lower is better

Loss Stability Variance analysis

Generalization Gap Val Loss - Train Loss

Model Fusion

LoRA/DoRA Adapter

Choose the trained adapter to fuse. The base model will be automatically detected from the adapter's configuration.

Fusion Progress

Ready for Fusion

Select a trained adapter, then click "Start Fusion" to begin. The base model will be automatically detected.

Progress 0%

Status

Initializing...

00:00

Elapsed Time

--:--

Est. Remaining

Current Job

Base Model: -

Adapter: -

Output: -

Model Quantization

Select Model to Quantize

Choose from local models or HuggingFace cache

Quantization Bits

Lower bits = smaller model, slightly lower quality

Group Size

Smaller groups = better quality, larger file

Quantized Models

No quantized models yet

Quantization Progress

Ready for Quantization

Select a model and click "Start Quantization" to begin

Progress 0%

Status

Initializing...

00:00

Elapsed Time

--:--

Est. Remaining

Current Job

Input: -

Output: -

Settings: -

Quantization Tips

4-bit vs 8-bit: 4-bit provides better compression (~75% size reduction) but slightly lower quality. 8-bit offers better quality with ~50% compression.

Group Size: Smaller groups (32) provide better quality but larger files. 64 is the recommended balance.

Performance: Quantized models are faster to load and use less memory, making them ideal for deployment.

Compatibility: Quantized models work with the same inference code as the original models.

Model Configuration

Base Model

Adapter Path (Optional)

Tip: You can select just an adapter to automatically load it with its base model!

System Prompt

Generation Parameters

Max Tokens

Context Window

Temperature

Top P

Rep. Penalty

Seed

Set seed for reproducible generation (same seed = same output)

Real-time Streaming

Shows text as it's generated

Generation Output

Model ready! Type a message below to start the conversation.