Annotation Export

RIME exports annotations as a flat Parquet file — one row per annotation. This is the primary format for downstream analysis, model benchmarking, and archiving.

File → Export Annotations

Columns

Identity

Column	Type	Description
`annotation_id`	`str`	Unique annotation ID
`session_id`	`str`	Session UUID
`session_name`	`str`	Human-readable session name
`subject_id`	`str`	Participant identifier
`rater`	`str`	Annotator name
`export_timestamp`	`str`	ISO 8601 UTC timestamp of this export

Annotation

Column	Type	Description
`lane`	`str`	Annotation lane (e.g. `FOG`, `Tasks`)
`label`	`str`	Label within the lane (e.g. `FOG`, `Walk`)
`event_type`	`str`	`interval` or `point`
`start_ms`	`float`	Onset in milliseconds from session start
`end_ms`	`float`	Offset in milliseconds (same as `start_ms` for points)
`duration_ms`	`float`	Duration in milliseconds

Provenance

Column	Type	Description
`source`	`str`	Where the annotation came from — see below
`confidence`	`float`	Current confidence score (1.0 for manual annotations)
`human_modified`	`bool`	`True` if a model annotation was subsequently edited by a rater
`origin_confidence`	`float \\| null`	Model's original confidence before any human adjustment
`origin_start_ms`	`float \\| null`	Model's original onset before any human adjustment
`origin_end_ms`	`float \\| null`	Model's original offset before any human adjustment
`ghost`	`bool`	`True` if the annotation was never accepted (excluded from exports by default)

The `source` field

source identifies the origin of every annotation:

Value	Meaning
`manual`	Created by a rater directly
`corrected`	Accepted from a model suggestion, then edited
`elan_import`	Imported from an ELAN `.eaf` file
`model:<name>`	Accepted from model `<name>` without modification

Provenance example

When a model produces an annotation and a rater later adjusts its boundaries:

Field	Value
`source`	`corrected`
`start_ms`	`12400` ← rater's adjusted onset
`end_ms`	`15800` ← rater's adjusted offset
`human_modified`	`True`
`origin_start_ms`	`12100` ← model's original onset
`origin_end_ms`	`16200` ← model's original offset
`origin_confidence`	`0.94` ← model's confidence at inference time

This lets you reconstruct both what the model predicted and what the rater accepted, in the same row.

Filtering ghost annotations

Ghost annotations (model suggestions not yet reviewed) are excluded from exports by default. To include them, enable Include unreviewed suggestions in the export dialog. Ghost rows have ghost = True and can be filtered out in analysis:

import pandas as pd

df = pd.read_parquet("session_annotations.parquet")
accepted = df[~df["ghost"]]

Loading in Python

import pandas as pd

df = pd.read_parquet("session_annotations.parquet")

# All accepted FOG episodes
fog = df[(df["lane"] == "FOG") & (~df["ghost"])]

# Model annotations that were subsequently edited
edited = df[df["human_modified"]]

# Compute onset correction (rater vs model)
edited["onset_correction_ms"] = edited["start_ms"] - edited["origin_start_ms"]