Available Models in Bios

1import bios
2import asyncio
3
4async def train_multiple_models(dataloader):
5    """
6    Train different expert models on the same data
7    Useful for comparing domain-specific performance
8    """
9    service_client = bios.ServiceClient()
10    
11    # Create clients for different models
12    models = {
13        "healthcare": await service_client.create_lora_training_client_async(
14            base_model="ultrasafe/usf-healthcare", rank=16
15        ),
16        "finance": await service_client.create_lora_training_client_async(
17            base_model="ultrasafe/usf-finance", rank=16
18        ),
19        "general": await service_client.create_lora_training_client_async(
20            base_model="ultrasafe/usf-mini", rank=8
21        )
22    }
23    
24    # Train all models in parallel
25    for step, batch in enumerate(dataloader):
26        # Submit training for all models
27        futures = {}
28        for name, client in models.items():
29            fwd_future = await client.forward_backward_async(batch, "cross_entropy")
30            opt_future = await client.optim_step_async()
31            futures[name] = (fwd_future, opt_future)
32        
33        # Collect results
34        for name, (fwd_future, opt_future) in futures.items():
35            fwd_result = await fwd_future
36            await opt_future
37            print(f"{name}: Step {step}, Loss = {fwd_result.loss:.4f}")
38    
39    # Save all models
40    paths = {}
41    for name, client in models.items():
42        path = client.save_state(name=f"{name}_final").result().path
43        paths[name] = path
44        print(f"Saved {name} model: {path}")
45    
46    return paths
47
48# Run multi-model training
49asyncio.run(train_multiple_models(training_data))