beautifulsoup-tasks-v1

Task dataset for a BeautifulSoup RL / eval environment, in the shape used by the Prime Intellect Environments Hub.

26 HTML-extraction tasks over three fixed documents. Each task gives the model the documents and an instruction; the answer is the list of rows left in result, compared exactly to a reference. Grading is deterministic — no LLM judge, no external API, no network.

Category	Tasks	Covers
selection	7	find_all, class filters, CSS selectors, attribute presence, select_one, href predicates
attributes	6	reading and casting data-*, meta content, entity decoding, missing attributes with a default, class lists
navigation	5	find_next_sibling, find_parent, direct children only (recursive=False), pairing a heading with its link
tables	4	header cells, full row extraction with type casting, filtering rows by class, aggregate over a column
malformed	4	unclosed <li>, a <div> inside a <p>, a stray </span>, an <img> with no alt

⚠ The determinism trap in this domain, and how it is handled

BeautifulSoup's result depends on the parser — and this was measured, not assumed. On the malformed document in this dataset, asking for the <p>'s text gives:

html.parser -> "Intro textnested blocktail text"   (keeps the <div> inside the <p>)
lxml        -> "Intro text"                        (closes the <p> first, per HTML5)

An answer key built with one parser and graded under another is simply wrong. Therefore:

• every task states BeautifulSoup(html, "html.parser") explicitly in its prompt;
• html.parser is used because it ships with CPython — no extra dependency can be missing or differently versioned in the sandbox;
• the malformed-HTML tasks are labelled as such, so the expected repair behaviour is the thing being tested rather than an accident.

Fields

Field	Description
task_id	stable id, e.g. bs-017
category	one of the five above
prompt	the natural-language instruction
documents	a short description of all three documents
doc_name	which document this task uses (catalogue, table, messy)
html	the full source of that document, so the task is self-contained
expected_output	{"rows": [[...], ...]} — the reference result

Verification

Every task is independently checked: it runs, is deterministic across two fresh parses, returns a non-empty list of JSON-safe primitives (which catches the common error of returning a Tag or NavigableString instead of extracted text), contains no accidental empty strings outside the tasks that legitimately allow one, and survives the serialisation round-trip exactly. All 26 pass on beautifulsoup4 4.15.0.

Builder and verifier: build_tasks.py.