danemarparceros/drupal7-dataset-api

Drupal 7 Dataset API

Drupal 7 reached end of life in early 2025, but plenty of sites still run on it, and the people maintaining them still need good answers. This dataset is meant to help with that: 7,952 examples about the Drupal 7 API, taken only from api.drupal.org. You can use it to fine-tune a model, to test how well a model knows Drupal 7, or as reference material for a retrieval (RAG) setup.

Everything is in English. It was put together by Daniel Ricardo Ramirez Marin as part of the Codicem project in Danemar Parceros.

What's inside

There are two configurations, so you can pick the one that fits what you're doing:

Config	Examples	What you get
reference	4,382	A question about an API symbol, with its exact signature and a short explanation as the answer
code	3,570	A coding task ("write the function X that...") with the real function body as the answer

In reference you'll find 4,010 functions, 247 constants, 91 classes and 34 globals. In code there are 2,501 functions, 739 hook implementations and 330 hook definition examples (the ones that come from *.api.php files).

Fields

Every record has the same fields in both configs:

Field	What it is
id	A stable identifier
drupal_version	Always "7"
source_type	drupal_api_reference or drupal_api_code
source_url	The api.drupal.org page the example comes from
instruction	The question or the task
context	In code: the description, parameters and return value. Empty in reference
response	The answer: signature plus explanation (reference), or the function source (code)
tags	The kind of example
created_at	When it was generated

A code example

{
  "instruction": "In Drupal 7, write the function `actions_do($action_ids, $object = NULL, ...)` (file includes/actions.inc) so that it does the following: Performs a given list of actions by executing their callback functions.",
  "context": "Given the IDs of actions to perform, this function finds out what the callback functions for the actions are by querying the database. ...\n\nParameters:\n- $action_ids: ...\n\nFile: includes/actions.inc",
  "response": "function actions_do($action_ids, $object = NULL, $context = NULL, $a1 = NULL, $a2 = NULL) {\n  // $stack tracks the number of recursive calls.\n  static $stack;\n  ..."
}

A reference example

{
  "instruction": "What does the comment_unpublish_action function do in Drupal 7?",
  "response": "function comment_unpublish_action($comment, $context = array()): Unpublishes a comment by setting its status to COMMENT_NOT_PUBLISHED and logs the action in watchdog."
}

How it was made

The two configs were built differently, so it's worth knowing which is which.

code didn't involve any model. It comes straight from the official Drupal 7 source code (core 7.104-dev). Each task is written from the function's docblock, and the answer is the function body copied exactly as it is, indentation included.

reference did involve models. For each api.drupal.org page, several language models proposed a question and an answer, and other models acted as judges, throwing out any pair that made things up or didn't copy the signature exactly. Those pairs were first written in Spanish and then machine-translated into English. The signature is never translated, and a translation was only accepted if every code identifier from the original (variables, function and constant names, quoted strings) came through untouched. So the questions and explanations in reference are AI-generated text, based only on what the page says.

What was left out

To keep the quality up, these were left out:

• simpletest scaffolding, which is mostly setUp and getInfo boilerplate (1,294)
• explanations that were empty or too short (35)
• signatures that don't appear word for word on the source page (5)
• anything that mentions another Drupal version (20)
• possible personal data (5)
• translations that didn't pass validation (28)
• class methods, because their source isn't a page that could be verified on api.drupal.org (413)

After the dataset was built

Once the dataset was fully created, I verified the data by hand, and then used it to fine-tune a Qwen 7B model.

License

The whole dataset is released under the GPL-2.0 or later, That's the license api.drupal.org itself declares for all of its code and documentation, and since this dataset is derived from that content, it keeps the same terms. Records don't carry a license field of their own because the license applies to the dataset as a whole.

If you use it, please keep the license and the reference to the source (source_url). Whether training a model on GPL content, or sharing that model, brings obligations of its own is still an open legal question, so if you plan to redistribute model weights, it's worth talking to someone who can advise you.

Drupal is a registered trademark of Dries Buytaert. This is an independent project and has no affiliation with, or endorsement from, the Drupal Association or the Drupal project.

Things to keep in mind

• The manual check was a verification, not a line-by-line review. The reference examples were checked by models during generation, and the translation was done by a model too, so some explanations may be a bit off. The code in code is official. I did verify that the fine-tuned Qwen 7B model uses Drupal 7 better than before, and it worked.
• It's Drupal 7 only, so it doesn't apply to Drupal 8 and later.
• Coverage isn't complete. There are no class methods and no tests, and in code functions with fewer than 2 lines of body or more than 6,000 characters are left out.

Citation

If this dataset is useful for your work, you can cite it like this:

@misc{danemarparceros/drupal7-dataset-api,
  title  = {Drupal 7 Dataset API},
  author = {Ramirez Marin, Daniel Ricardo},
  year   = {2026},
  note   = {Dataset derived from api.drupal.org, GPL-2.0-or-later license}
}