Premove ITN Python API

The public contextual API has one implementation: PremoveITN. The current package release is v0.3.0. It adds structured and contextual views while retaining the frozen v0.2.0 model artifact. The model scores and decoder output remain one inference result; temporal enrichment is deterministic post-processing over that result.

Load the model

from premove_itn import PremoveITN

itn = PremoveITN.from_pretrained()

from_pretrained() downloads the pinned public model on first use and uses the Hugging Face cache on later loads. Create one instance and reuse it. Model initialization is expensive.

Parameters

PremoveITN.from_pretrained(
    model_id="premove-ai/premove-itn",
    *,
    revision="e42a6ad5f58d3fde9cb6cf1f81f7fe40b9d99526",
    device="auto",
    context=None,
)

model_id accepts the public model ID or a local inference-artifact directory.
revision accepts the pinned model commit or the verified v0.2.0 model release tag.
device accepts auto, cpu, mps, or cuda.
context accepts an optional default NormalizationContext for later calls.

The loader verifies release metadata and the model-file digest before inference. An arbitrary Hub repository is not accepted. A local artifact must match the frozen release contract. device="auto" selects CUDA when available, then Apple MPS, then CPU. CUDA is an API option, but it is not a validated v0.3.0 platform claim.

Update an existing installation

Upgrade the package with the normal Python package manager:

python -m pip install --upgrade premove-itn

After the upgrade, PremoveITN.from_pretrained() uses the current package default and downloads the matching pinned model snapshot if it is not cached. Hugging Face keeps snapshots by revision, so an older cached release is not overwritten. There is no silent background package update.

Normalize a transcript

result = itn.normalize("call me at four thirty")
print(result)
# call me at 04:30

normalize() accepts one string and returns one string. It raises TypeError for a non-string input. Empty and whitespace-only strings are returned unchanged. If the text produces candidates, inputs longer than 512 DeBERTa encoder tokens are rejected instead of being truncated. Text with no candidates returns unchanged before tokenization.

Inspect structured normalization

Use normalize_structured() when a downstream system needs the exact edits selected by ITN:

result = itn.normalize_structured("pay twenty dollars")

print(result.text)
# pay $20

span = result.spans[0]
print(span.source_text, span.normalized_text)
# twenty dollars $20

The result is an immutable NormalizationResult with:

text: readable normalized text.
resolved_text: the machine-resolved text view.
spans: selected edits as immutable NormalizedSpan values.

Each span retains exact half-open source and normalized character offsets, source and normalized text, every contributing SpanKind, and an optional resolved_value. These invariants hold for every selected span:

source[span.source_start : span.source_end] == span.source_text
result.text[span.normalized_start : span.normalized_end] == span.normalized_text

The decoder renders the selected path once. The normalized text and offsets come from that same render; the API does not recover alignment with a later diff or a second inference pass. The initial contextual resolver supports today, tomorrow, yesterday, day after tomorrow, and day before yesterday. It keeps these expressions in readable text and adds ISO calendar values to resolved_value and resolved_text when reference_datetime is available. Without a reference datetime, the expressions remain annotated but unresolved. Selected named-month DATE spans without a year are also resolved with the reference year. Explicit years always win. Invalid calendar dates remain unresolved. Weekday-qualified named dates are resolved only when the stated weekday agrees with the resulting calendar date. A missing year uses the reference year; an explicit year can resolve without context. Contradictory weekdays remain unresolved. Selected numeric DATE spans are resolved when their interpretation is structurally unique. Ambiguous dates require DateOrder; otherwise they stay unresolved. Numeric dates without a year require the reference year. The resolver accepts slash, dash, dot, and space-separated fields without adding format-specific context fields. When the year is omitted, DateOrder uses the relative position of day and month; all six orders therefore collapse to either day-month or month-day interpretation. Several temporal expressions can be resolved in one result. The structured view preserves every selected decoder span and every compatible contextual DATE span, while resolved_text renders all resolved values in normalized-text order. The three public views share one inference result; resolving temporal metadata does not trigger a second model call.

Supply normalization context

Create an immutable context from explicit caller-owned facts:

from datetime import datetime

from premove_itn import DateOrder, NormalizationContext, PremoveITN

context = NormalizationContext(
    reference_datetime=datetime(2026, 9, 20, 12, 0),
    timezone="Asia/Kolkata",
    locale="en-IN",
    date_order=DateOrder.DMY,
)

itn = PremoveITN.from_pretrained(context=context)

The instance context is a default. A per-call context replaces it completely:

result = itn.normalize_structured(
    "show my meetings tomorrow",
    context=NormalizationContext(
        reference_datetime=datetime(2026, 10, 1, 9, 0),
        timezone="America/New_York",
        locale="en-US",
        date_order=DateOrder.MDY,
    ),
)

Contexts are not merged field by field. Premove ITN does not discover the current time, timezone, locale, or date order. The resolver consumes only the explicit context supplied by the caller. Pass an empty NormalizationContext() to replace an instance default with no contextual facts for one call. Passing context=None uses the instance default. DateOrder describes only the positional order of day, month, and year. It supports all six permutations: DMY, DYM, MDY, MYD, YDM, and YMD. It does not encode separators or surface formats. For example, 30/09/2026, 30-09-2026, and 30 09 2026 all use DateOrder.DMY. Named-month dates do not need a date order when their fields are already unambiguous.

Return the resolved text view

normalize_resolved() is the string-only view of the same internal result:

resolved = itn.normalize_resolved("pay twenty dollars")
print(resolved)
# pay $20

It does not run a separate inference or reconstruction path. Its output will differ from normalize() when a supported contextual expression has a deterministic value to render:

resolved = itn.normalize_resolved(
    "call me tomorrow",
    context=NormalizationContext(
        reference_datetime=datetime(2026, 9, 20, 12, 0),
    ),
)
print(resolved)
# call me 2026-09-21

Reuse one instance

texts = [
    "the total is twenty dollars",
    "email support at example dot com",
    "the room code is one oh five",
]

for text in texts:
    print(itn.normalize(text))

For model lifecycle and supported environments, see Deployment. For a streaming shell workflow, see the CLI.

Overview

Integrate

Understand

Run in production

Learn

Internals

Premove ITN Python API

Load the model

Parameters

Update an existing installation

Normalize a transcript

Inspect structured normalization

Supply normalization context

Return the resolved text view

Reuse one instance

​Load the model

​Parameters

​Update an existing installation

​Normalize a transcript

​Inspect structured normalization

​Supply normalization context

​Return the resolved text view

​Reuse one instance

Load the model

Parameters

Update an existing installation

Normalize a transcript

Inspect structured normalization

Supply normalization context

Return the resolved text view

Reuse one instance