Task Query API

Training-role index entries and non-withheld activity phrases yield distinct (text, level) queries. Their target/referring-code sets drive task cross-entropy; held-out text stays sealed.

Req 11's task queries (spec 4.1(i)), built once from a validated supervision bundle.

A query is a distinct (text, level) pair with a target set T of codes at that level and a set N of forced negatives. Two bundle members give them:

  • The role table's training entries: each is a query at its code's level, 6, with T its code and N empty.
  • The redirection table's activity phrases. A phrase row's destinations are its named codes that are not lineal to its referencing code, so a row whose named codes are all lineal stays text only (Req 8), and a withheld row carries no phrase. Rows group by (phrase, destination level): T is the group's destinations at that level, and N its rows' referencing codes, whatever their level, so a cross-reference query always scores its referencing code (Req 8(b)).

An index entry and a phrase with the same exact text at level 6 are one query, whose T and N are the unions; at another level the two stay apart. A query whose T and N overlap is refused: no code can be both where an activity is classified and the code that sends it elsewhere.

Only training text is read: the role table's training rows, and the phrases of the redirection rows that are not withheld, without the rows' texts.

TaskQuery dataclass

One task query (spec 4.1(i)).

Attributes:

Name Type Description
text str

The query's text: an index entry's or an activity phrase, exactly as stored.
level int

The level of every target, 2-6.
targets Tuple[str, ...]

T, the codes the text is classified in, sorted.
negatives Tuple[str, ...]

N, the forced negatives: the referencing codes of the phrase rows that send the text away, sorted.

build_task_queries(bundle)

Every task query of the bundle, sorted by level, then text (spec 4.1(i)).

Parameters:

Name Type Description Default
bundle ValidatedSupervisionBundle

A validated supervision bundle, whose index_roles and redirections members give the queries.

 required

Returns:

Type Description
List[TaskQuery]

One query per distinct (text, level) pair, with its targets and forced negatives sorted.

Raises:

Type Description
ValueError

If a query's targets and forced negatives overlap.