Full Control

For full control over training and inference, use the lower-level Python classes.

Custom Training

import pmetal

# Configure training components
lora_config = pmetal.LoraConfig(r=16, alpha=32.0)
training_config = pmetal.TrainingConfig(
    learning_rate=2e-4,
    num_epochs=3,
    batch_size=4,
    max_seq_len=2048,
)

# Create trainer
trainer = pmetal.Trainer(
    model_id="Qwen/Qwen3-0.6B",
    lora_config=lora_config,
    training_config=training_config,
    dataset_path="train.jsonl",
)

# Add callbacks
trainer.add_callback(pmetal.ProgressCallback())
trainer.add_callback(pmetal.LoggingCallback())

# Train
result = trainer.train()

Model Loading

# Load model
model = pmetal.Model.load("Qwen/Qwen3-0.6B")

# Generate
output = model.generate("Hello world", temperature=0.7)
print(output)

Configuration Classes

Class	Description
pmetal.LoraConfig	LoRA rank, alpha, target modules
pmetal.TrainingConfig	Learning rate, epochs, batch size, scheduler
pmetal.GenerationConfig	Temperature, top-k, top-p, max tokens
pmetal.DataLoaderConfig	Dataset format, sequence packing

Callbacks

Callback	Description
pmetal.ProgressCallback()	Progress bar display
pmetal.LoggingCallback()	Console logging
pmetal.MetricsJsonCallback(path)	JSONL metrics file
Custom	Subclass with on_step_end(step, loss)

Hub Operations

# Download a model
pmetal.download_model("Qwen/Qwen3-0.6B")

# Download a specific file
pmetal.download_file("Qwen/Qwen3-0.6B", "config.json")

Tokenizer

tokenizer = pmetal.Tokenizer("Qwen/Qwen3-0.6B")
tokens = tokenizer.encode("Hello world")
text = tokenizer.decode(tokens)

See Also

• Quick Start — Simplified API
• Rust SDK — Rust equivalent