A community, unofficial, open source toolkit for starting fast with Adaption's Adaptive Data and AutoScientist.
This is an independent, community project. It is not affiliated with, sponsored by, or endorsed by Adaption Labs. "Adaption", "Adaptive Data", "AutoScientist", and any related names are trademarks of their respective owners and are used here only to describe what this toolkit helps you do. Always treat the official Adaption documentation and API as the source of truth. Nothing here is official support.
For the platform itself, and the team behind it, go to the source:
adaption-devkit is a small kit that gets a beginner or a student from a raw CSV to a published, adapted model without burning credits on avoidable mistakes. It is built around the Adaptive Data lifecycle:
flowchart LR
A[ingest] --> B[adapt] --> C[evaluate] --> D[export] --> E[publish]
classDef stage fill:#0b7285,stroke:#08505c,color:#ffffff;
class A,B,C,D,E stage;
Read the official docs to learn the platform, then reach for this kit to move faster and skip the small mistakes that cost real time and credits the first time around:
- A preflight dataset linter that catches the always on deduplication collapse before you spend credits. Near duplicate rows can shrink your dataset to a fraction of its size after you have already paid. The linter warns first.
- A publish helper, because the official publish endpoint currently returns HTTP 501. The helper packages your release so you can ship to Hugging Face and Kaggle anyway.
- Estimate first run helpers so you see the credit cost before each run.
- Decision guides for column mapping and for recipes and brand controls, so you pick the right levers for your domain instead of guessing.
- Ready templates and cookbook notebooks so your first run works on the first try.
adaption-kit doctor- check your setup in one shot: Python, the SDK, your env vars, the host.adaption-kit lint- preflight linter for your dataset, run it before you pay.adaption-kit suggest- look at your file and recommend the column mapping to use.adaption-kit convert- convert between CSV, JSONL, and Parquet.adaption-kit estimate- quote credits and time before a run.adaption-kit run- estimate first wrapper around an adaptation run.adaption-kit publish- publish helper for the 501 endpoint.adaption-kit card- generate dataset and model cards.adaption-kit cover- generate a cover image for your release.- Guides, runnable cookbook notebooks, and ready to edit templates.
git clone https://github.com/A1VARA5/adaption-devkit.git
cd adaption-devkit
pip install -e .Optional extras:
# the SDK-backed run and publish helpers
pip install -e ".[sdk]"
# everything for the cookbook notebooks
pip install -e ".[notebooks]"If an extra is not installed, the matching command tells you what to add. The core linter and the guides work with no extras at all.
Configure access through environment variables. Never hardcode your key.
export ADAPTION_BASE_URL="https://api.prod.adaptionlabs.ai"
export ADAPTION_API_KEY="your-key-here"On Windows PowerShell:
$env:ADAPTION_BASE_URL = "https://api.prod.adaptionlabs.ai"
$env:ADAPTION_API_KEY = "your-key-here"Then lint your data before you spend anything:
adaption-kit lint data.csvThe linter reports the columns it sees, flags near duplicate rows that the
deduplication pass would collapse, and tells you whether your column mapping looks
right. Fix the warnings, then move on to estimate and run.
| Path | What is in it |
|---|---|
adaption_kit/ |
the Python package and the adaption-kit CLI (doctor, lint, suggest, convert, estimate, run, publish, card, cover) |
guides/ |
quickstart, gotchas, column-mapping, recipes-and-controls, release-checklist |
cookbook/ |
runnable notebooks that walk the full lifecycle |
templates/ |
dataset schemas, dataset and model cards, a cover, and Kaggle metadata |
graphics/ |
the diagrams embedded below, as Mermaid in Markdown |
tests/ |
the pytest suite, run with pytest |
examples/ |
full walkthroughs you can copy and paste, marketing and question and answer |
MAINTENANCE.md |
what is verified versus correct by construction, and how current it is |
LICENSE |
Apache-2.0 |
CONTRIBUTING.md |
how to contribute and the quality bar |
CODE_OF_CONDUCT.md |
Contributor Covenant v2.1 |
flowchart LR
A[ingest<br/>upload your data] --> B[adapt<br/>run a recipe]
B --> C[evaluate<br/>measure improvement_percent]
C --> D[export<br/>download the result]
D --> E[publish<br/>release to HF and Kaggle]
L([adaption-kit lint]) -.-> A
M([adaption-kit estimate]) -.-> B
P([adaption-kit publish]) -.-> E
classDef stage fill:#0b7285,stroke:#08505c,color:#ffffff;
classDef helper fill:#f1f3f5,stroke:#adb5bd,color:#212529;
class A,B,C,D,E stage;
class L,M,P helper;
See graphics/command-router.md. A quick route from a standing start to a published release.
flowchart TD
S([new here? start here]) --> Doc[adaption-kit doctor]
Doc --> Q1{know your column mapping?}
Q1 -- no --> Sug[adaption-kit suggest]
Q1 -- yes --> Lint[adaption-kit lint]
Sug --> Lint
Lint --> Est[adaption-kit estimate]
Est --> Run[adaption-kit run<br/>pilot first]
Run --> Q2{number look good?}
Q2 -- no --> Tune[change one lever]
Tune --> Est
Q2 -- yes --> Full[adaption-kit run<br/>full corpus]
Full --> Pub[adaption-kit publish]
classDef cmd fill:#0b7285,stroke:#08505c,color:#ffffff;
classDef gate fill:#e7f5ff,stroke:#1c7ed6,color:#0b4884;
class Doc,Sug,Lint,Est,Run,Tune,Full,Pub cmd;
class Q1,Q2 gate;
See graphics/column-mapping-decision.md.
flowchart TD
Start([What is in your dataset?]) --> Q1{Prompts and answers together?}
Q1 -- Yes --> Both[Map both<br/>prompt + completion]
Q1 -- No --> Q2{Which one do you have?}
Q2 -- Answers only --> Completion[Map as completion]
Q2 -- Prompts only --> Prompt[Map as prompt]
Q2 -- Reference passages --> Context[Map as context]
classDef q fill:#e7f5ff,stroke:#1c7ed6,color:#0b4884;
classDef out fill:#0b7285,stroke:#08505c,color:#ffffff;
class Q1,Q2 q;
class Both,Completion,Prompt,Context out;
See graphics/recipe-matrix.md for the full table.
See graphics/credit-safe-run.md. Estimate, pilot small, read the number, change one lever, then scale.
flowchart TD
A([lint your dataset]) --> B[estimate the run]
B --> C[pilot on a small max_rows slice]
C --> D{improvement_percent}
D -- low --> E[change one lever<br/>a recipe or a brand control]
E --> B
D -- good --> F[run the full corpus]
F --> G([publish to Hugging Face and Kaggle])
classDef step fill:#0b7285,stroke:#08505c,color:#ffffff;
classDef gate fill:#e7f5ff,stroke:#1c7ed6,color:#0b4884;
class A,B,C,E,F,G step;
class D gate;
See graphics/publish-flow.md. The publish endpoint returns 501, so you release by hand.
flowchart LR
A[run done] --> B[download the adapted dataset]
B --> C[write a dataset card]
C --> D[render a cover image]
D --> E[push to Hugging Face]
D --> F[push to Kaggle]
classDef step fill:#0b7285,stroke:#08505c,color:#ffffff;
class A,B,C,D,E,F step;
Contributions are welcome and the quality bar is simple: only verified, correct
content, and anything untested is marked as such. Read
CONTRIBUTING.md and the
CODE_OF_CONDUCT.md before you open a pull request.
Apache-2.0. Copyright 2026 Aivaras Navardauskas.
Author: Aivaras Navardauskas (MANIFESTA). GitHub: A1VARA5.