Module captioning 

Three-level hierarchical caption generation pipeline.

SensorLM’s key insight is that paired (sensor, text) training data can be generated automatically from unlabelled wearable recordings, eliminating the need for human annotation at scale.

Caption levels

Level	Module	Description	Token budget
1 – Statistical	statistical	Mean/max/min/std per channel	512
2 – Structural	structural	Trends & anomaly events	512
3 – Semantic	semantic	Activities, sleep, mood	256–1024

Combination keys

The training pipeline selects one of eight caption variants for each batch:

low_level_caption             → level 1 only
middle_level_caption          → level 2 only
high_level_summary_caption    → level 3 only (short)
high_level_all_caption        → level 3 (full)
middle_low_level_caption      → levels 2 + 1
high_low_level_caption        → levels 3 + 1
high_middle_level_caption     → levels 3 + 2
high_middle_low_level_caption → levels 3 + 2 + 1

Modules

semantic

Level-3 (semantic) caption generation.

statistical

Level-1 (statistical) caption generation.

structural

Level-2 (structural) caption generation.

templates

Text templates for all three captioning levels.

Structs

CaptionContext

All contextual information needed to produce a full multi-level caption.

Functions

generate_caption

Generate the caption text for the requested CaptionKey.