Skip to main content
Adaption is currently in Beta and may not be stable. Users may also encounter slower outputs due to vLLM cold-start times.
Adaption lets you asynchronously finetune your models as they interact in the real world, ensuring your agents continually improve over time. Use trigger_training() to start a finetuning job on your collected memories. Then use monitor_training() to check the status of your job and look for "lora_ready": true in the response.

​
Trigger Training

import requests

def trigger_training():
    response = requests.post(
        "https://rkdune--symmetry.modal.run/adaption/train",
        headers={"Authorization": "Bearer YOUR_ASYMMETRIC_API_KEY"},
        json={
            "memory_group": "darwin_agent",
            "lora_name": "darwin_adapter",
            "force": True
        }
    )
    print(response.json())

trigger_training()

​
Monitor Training

def monitor_training():
    response = requests.get(
        "https://rkdune--symmetry.modal.run/adaption/status",
        headers={"Authorization": "Bearer YOUR_ASYMMETRIC_API_KEY"},
        params={
            "memory_group": "darwin_agent",
            "lora_name": "darwin_adapter"
        }
    )
    print(response.json())  # Look for "lora_ready": true

monitor_training()

​
Run Inference with Your Finetuned Adapter

Once training is complete, use your finetuned adapter at inference time by setting adaption_inference: True and specifying your lora_name. 
from openai import OpenAI

client = OpenAI(
    base_url="https://rkdune--symmetry.modal.run/v1/",
    api_key="YOUR_ASYMMETRIC_API_KEY",
)

completion = client.chat.completions.create(
    model="asymmetric/Qwen3-8B",  # use any model
    messages=[{"role": "user", "content": "How old was Darwin when he set off on his voyages?"}],
    stream=False,
    extra_body={
        "nightly": ["darwin_agent", "This agent helps collect historical record about Darwin's expeditions."],
        "finetune_thresh": 1,
        "min_finetune_group": 1,
        "lora_name": "darwin_adapter",
        "adaption_inference": True,
    }
)

print(completion.choices[0].message.content)