smilegeng/flock_demo_glucose

Flock Demo Glucose Time-Series — TsFile format

This repository is a conversion to TsFile format of the Hugging Face dataset random-sequence/flock-demo-time-series-prediction-sections.

⚠️ This is a demo dataset: simulated patient glucose / insulin / carbohydrate / activity monitoring time series, not real clinical data.

• Original dataset: random-sequence/flock-demo-time-series-prediction-sections
• License: the original dataset declares no license (this conversion follows suit and adds none).
• Original README: only field definitions (YAML frontmatter), no body description. The data-semantics description below comes from inspecting the actual data.

Data content

Simulated continuous monitoring records for 10 patients, one sample every 5 minutes. The original dataset is split by time into 8 sections (section_01 … section_08, each a Hugging Face config); the 8 sections are contiguous in time with no overlap or gaps:

section	rows	patients	time range
section_01	250	10	2024-01-01 00:00 → 2024-01-01 20:45
section_02	250	10	2024-01-01 20:50 → 2024-01-02 17:35
section_03	250	10	2024-01-02 17:40 → 2024-01-03 14:25
section_04	250	10	2024-01-03 14:30 → 2024-01-04 11:15
section_05	250	10	2024-01-04 11:20 → 2024-01-05 08:05
section_06	250	10	2024-01-05 08:10 → 2024-01-06 04:55
section_07	250	10	2024-01-06 05:00 → 2024-01-07 01:45
section_08	250	10	2024-01-07 01:50 → 2024-01-07 22:35

Within each section: the global timestamp is unique and strictly increasing, with the 10 patients interleaved; per patient, each patient's own records are also monotonic in time with no duplicates.

Files (one TsFile per section)

section_01.tsfile … section_08.tsfile, one-to-one with the original dataset's 8 configs, not merged.

TsFile structure

• TAG (device dimension) = patient_id: each patient is a device. Within each patient the timestamp is monotonic with no duplicates.
• Time: parsed from the original timestamp (2024-01-01 00:00:00 text, 5-minute interval) into INT64 milliseconds.
• FIELD:
  • glucose_mg_dl (DOUBLE, blood glucose mg/dL)
  • insulin_units (DOUBLE, insulin units)
  • carbs_grams (DOUBLE, carbohydrate grams)
  • activity_level (STRING, activity level: rest / light / moderate / intense)

What changed vs the original dataset

Item	Original	This TsFile version	Notes
Format	Parquet (8 configs)	TsFile (8 .tsfile)	one section per TsFile, preserving the original split
timestamp column	string (2024-01-01 00:00:00)	parsed into Time (INT64 ms) and the original text column dropped	time info fully preserved; only the format changes from string to millisecond integer
patient_id	int64 regular field	declared as TAG (device), stored as STRING in TsFile	TAG columns in the TsFile table model are unified to STRING; the integer value is preserved verbatim (1 → "1")
other columns	—	kept as-is	glucose_mg_dl / insulin_units / carbs_grams → DOUBLE, activity_level → STRING
row order	by global time	sorted by (patient_id, Time)	guarantees monotonic time within each device, as required for TsFile writes

The only dropped column: the original timestamp text column (its info is folded into Time). Apart from that, no columns are dropped and no values are modified.

Reading example

from tsfile import TsFileReader
reader = TsFileReader("section_01.tsfile")
for name, schema in reader.get_all_table_schemas().items():
    print(name, [c.get_column_name() for c in schema.get_columns()])