Adaptive Data App Walkthrough

Before you start

Sign in at adaptionlabs.ai/app. You do not need an API key for this app walkthrough.

Adaptive Data can build:

• Instruction datasets, which pair prompts with expected completions for supervised fine-tuning.
• Preference pairs, which compare two completions and identify the preferred one.

Step 1 | Bring in your data

Create a dataset in the app from one of these sources:

• Local file: CSV, JSON, JSONL, Parquet, PDF, DOCX, PPTX, XLSX, HTML, ZIP, or TXT.
• Hugging Face: enter a dataset URL and select the files to import. If the dataset is gated, request access on Hugging Face before importing it.
• Kaggle: enter the dataset page URL.

Document-style sources such as PDFs, scans, slide decks, and email threads trigger an additional extraction and splitting flow. See Processing unstructured documents.

Adaptive Data does not currently support non-text tabular data such as financial time series.

Step 2 | Expand the dataset

In the Expansion step, optionally increase the size of the dataset:

• Translate creates translated versions of existing rows. The app supports up to 242 languages; see the supported language list.
• Localize adapts rows for both a language and a country, including locale-specific wording rather than direct translation alone.

Skip expansion when you need to preserve the current dataset size.

Step 3 | Map your columns

Assign each source column a role:

• Prompt: the question, instruction, or task in each row.
• Context: background material needed to answer the prompt. You can select multiple context columns, including images.
• Completion: the expected response.
• Chat: a multi-turn conversation or dialogue history.

Every run needs at least a prompt or a completion. Adaptive Data can generate the missing role.

• No prompt column: select I don’t have prompt, then either let Adaption create one per row or select Write universal prompt.
• No completion column: select I don’t have a completion column to generate one.

For example, a document-extraction dataset could use document_body as context and this universal prompt:

Extract the customer name, contract start date, renewal date, and total contract value. Return valid JSON with the keys customer_name, start_date, renewal_date, and contract_value.

Make sure every column needed to answer the prompt is mapped as context. The column-selection guide includes detailed mapping patterns for image columns.

Step 4 | Evaluate data quality

The Data Evaluation step reports:

• A letter grade and quality score from 0 to 10.
• A percentile relative to other datasets in the detected domain.
• Detected domain, language, and tone.

Use this baseline to compare the source dataset with the adapted result after the run.

Step 5 | Choose recipes

Select only the recipes that fit your goal:

• Prompt Deduplication removes duplicate prompts so repeated examples do not dominate the dataset.
• Prompt Rephrasing rewrites prompts for clarity and alignment with the task.
• Reasoning Traces adds step-by-step reasoning alongside final outputs for supported workflows.
• Hallucination Mitigation applies grounding and verification to reduce unsupported completions.
• Freeform Brand Voice steers generated data toward a particular personality, tone, and style.

If prompts must remain byte-for-byte unchanged, disable Prompt Rephrasing and Metadata Injection.

Step 6 | Configure Blueprint

If you answer Yes to Do you want to add brand guidelines or global constraints?, the app opens Blueprint and its brand controls.

Enter a system-level instruction directly, or upload brand guidelines, a system prompt, or policies for safety, tone, and length. For example:

Write concise customer-support answers in British English. Use a warm, direct tone. Never recommend third-party products, and do not make claims that are not supported by the supplied context.

Blueprint also provides structured controls:

• Length: Detailed, Extensive, Minimal, or Concise.
• Safety Classification: Sexual, Self-harm, Hate, Harassment, and Violence.
• Hallucination mitigation: reduces unsupported content in generated completions.

These controls apply to every adapted row, so use them for true global requirements rather than row-specific guidance.

Step 7 | Review and launch

The Summary screen shows:

• A Reduce Rows slider for processing fewer rows.
• Your available credits.
• The exact credit cost of the configured job.

Review the estimate, then launch the run. The dashboard can show these states:

• Job launched: submitted or queued.
• Running: actively processing.
• Job completed: results and a quality score are available.
• Job failed: processing stopped before completion.

Large imports can spend several hours in the queue. If a job appears stuck:

1. Compare the elapsed time with the displayed estimate and allow for queue time.
2. Confirm the dataset uses supported prompt/completion columns or OpenAI-style chat data.
3. Retry a failed job once, especially if it reached 100% before failing.
4. If it remains stuck or fails again, contact support and include the dataset ID.

Step 8 | Review results

When processing finishes, open the dataset and select the View tab. Each row shows source and adapted data side by side, including original and enhanced prompts, original and enhanced completions, and inline context.

Compare the before-and-after:

• Quality score, from 0 to 10.
• Letter grade.
• Domain percentile, from 0 to 100.
• Individual source and adapted rows.

Row count can decrease because prompt deduplication removes duplicate examples.

Common issues

• Job launched for hours: queue time may not be included in the estimate. Wait beyond the displayed estimate before escalating.
• Pending Hugging Face or Kaggle image import: if a multimodal import does not progress, contact support with the dataset ID.
• Processing failed for a large JSONL file: retry once; if it fails again, contact support with the dataset ID.
• Dataset rejected for format: reshape it into prompt/completion columns or an OpenAI-style chat column, then repeat column mapping.
• Prompts changed unexpectedly: disable Prompt Rephrasing and Metadata Injection for strict prompt fidelity.

Next steps

• Work with image columns
• Configure Adaptive Data in detail
• Process unstructured documents
• Use the API reference when you are ready to automate the same workflow.