Reasoning traces

Reasoning traces are part of Adaptive Data’s recipe library—alongside options such as deduplication and prompt rephrasing—so you can shape how models reason over your data, not only the final string.

Traces expose an intermediate reasoning path with each adapted completion. That supports auditability (you can inspect how an answer was derived), distillation, and supervised fine-tuning that teaches models to show structured reasoning before you strip traces for production.

In the SDK this is a recipe toggle under recipe_specification, not a brand_controls field.

Exact output shape (how the trace relates to the completion) depends on the API and recipe version. Inspect downloaded results and the API Reference for RecipeSpecification and response types.

Configure a run

run = client.datasets.run(
    dataset_id,
    column_mapping={
        "prompt": "instruction",
        "completion": "response",
    },
    recipe_specification={
        "recipes": {
            "reasoning_traces": True,
        },
    },
)

Full example

import os
import time

from adaption import Adaption, DatasetTimeout

client = Adaption(api_key=os.environ["ADAPTION_API_KEY"])
dataset_id = os.environ.get("ADAPTION_DATASET_ID")

if not dataset_id:
    result = client.datasets.upload_file("training_data.csv")
    dataset_id = result.dataset_id
    while True:
        st = client.datasets.get_status(dataset_id)
        if st.row_count is not None:
            break
        time.sleep(2)

run = client.datasets.run(
    dataset_id,
    column_mapping={"prompt": "instruction", "completion": "response"},
    recipe_specification={"recipes": {"reasoning_traces": True}},
)
print(f"Run started: {run.run_id}, ~{run.estimated_credits_consumed} credits")

try:
    final = client.datasets.wait_for_completion(dataset_id, timeout=3600)
    print(f"Finished: {final.status}")
    if final.error:
        raise RuntimeError(final.error.message)
except DatasetTimeout:
    print("Timed out — check status manually")

url = client.datasets.download(dataset_id)
print(f"Download: {url}")

Combining with other recipes

RecipeSpecificationRecipes includes additional toggles (for example deduplication, preference_pairs, prompt_rephrase). Set multiple keys under recipes in one run; omitted keys keep backend defaults.

When to use it

Use reasoning traces when interpretable intermediate steps matter: debugging adaptation quality, building chain-of-thought–style training data, or meeting internal review requirements. Omit them when you only need short final answers and want to minimize tokens and cost.