ProVoice-proactivity/proactivity_preference_dataset

ProVoice study 1 — driver state, vehicle context and preferred Level of Autonomy

Driving-simulator data from the population data collection of the ProVoice / ProActivity project (CARLA 0.10): 12 drivers × 2 sessions, ~20 Hz multimodal driver-state and vehicle frames, and 1,446 driver-assigned Level-of-Autonomy (LoA) labels stating how autonomously an in-vehicle assistant should act on a given task. Drivers were prompted every 20 s about two randomly drawn in-vehicle tasks and marked, for each, the LoA(s) they would accept (0 = do nothing … 4 = act autonomously).

Config	What it is	Use it for
labels	one row per prompt: the driver's answer plus the window it refers to	the ground truth; join key for everything else
frames_preprocessed	every logged frame with the heart-rate channel repaired offline	driver-state modelling (this is what the released models were trained on)
frames_raw	the same frames exactly as logged live	provenance; comparing the live vs. offline HR filter
labeled	frames_preprocessed joined to labels, frames duplicated once per label	one-file training input; read §C before counting rows

data/labels.csv is the same table as the labels config in the CSV form scripts/build_loa_dataset.py reads (the Hub loads one file format per repo). calibration/ (not a config) holds each driver's 60 s calibration baseline and per-tick calibration log — inputs of the heart-rate repair, shipped so that frames_preprocessed can be regenerated from frames_raw.

from datasets import load_dataset
labels = load_dataset("danipulidoe/proactivity_preference_dataset", "labels", split="train")
frames = load_dataset("danipulidoe/proactivity_preference_dataset", "frames_preprocessed", split="train")

Reproducing the pipeline

Files are in the exact formats the code reads (JSONL frames, CSV labels), so the project's documented commands run on them unchanged (code: https://github.com/DaniPulidoE/ProActivity_Personalization). Code version used to build and verify this release: d961ce4bd191.

hf download danipulidoe/proactivity_preference_dataset --repo-type dataset --local-dir provoice_study1

# 1. frames_raw + calibration  ->  frames_preprocessed   (heart-rate repair)
python data_preprocessing/heart_rate_preprocessing.py \
    --in-data  provoice_study1/data/frames_raw.jsonl \
    --calib-dir provoice_study1/calibration \
    --out-data provoice_study1/data/frames_preprocessed.jsonl --no-write-calibration

# 2. frames_preprocessed + labels  ->  labeled            (window alignment)
python scripts/build_loa_dataset.py \
    --raw    provoice_study1/data/frames_preprocessed.jsonl \
    --labels provoice_study1/data/labels.csv \
    --out-jsonl provoice_study1/data/labeled.jsonl --out-fcd fcd_out.csv

# 3. labeled  ->  population model
python -m ProVoice.models.train_XLSTM --in provoice_study1/data/labeled.jsonl \
    --out trained_models/state_xlstm.pt --loss corn

Steps 1 and 2 were re-run on the released files at build time and reproduce the released frames_preprocessed and labeled row for row (see "Verification" at the end). The columns listed under "excluded" in the schema below are absent from the released files; none of them is read by any step.

File	SHA-256
calibration/calibration_001.json	33ca2fa2ea4965b10933327ddb8760cd646f45697df403f21a61637947c8067c
calibration/calibration_002.json	b5f694ea0d55a862d8c8d34d332276c8ef0408308ce975fbd4b1c79fe59ff3b2
calibration/calibration_003.json	4b4d0b9e9be04d68d6e3d81ebe513ac572eccc89ac750199b31a35a21fc2831a
calibration/calibration_004.json	acca11fc08a959e3dad41f051f13d6964b37bd991d9c66ce3f25300a02537fba
calibration/calibration_005.json	7d63dc5f40a04520f7078dc690b11f122792ec3bb7ed8ad7664dbb6c1e6483e5
calibration/calibration_006.json	2c7b2eb86bdc45a44b9ade7c0e791f0eaa49b40b02e20b03e8e0b7649a55e67c
calibration/calibration_007.json	cfd1f551b75a3c4550bcfccc502752ab682a38db91d2f18f70b0a7cc2ed737b5
calibration/calibration_008.json	0d362cf1dd78289eb97c10bcb2bf9cef139e7c8f6d5ce81418291e26df082a7f
calibration/calibration_009.json	535e2a3ab671a8a319bee4bc1140f2b4a30bbe6fc9e7cf7ecc05bbeaf65cbe2e
calibration/calibration_010.json	11421b92d2672a7afbe3606b9a26cdfe2df91a3890b9942800b1c921f3bf1d20
calibration/calibration_011.json	70315e1020313369550f5a56d75b76b75f4de19ef7315362a7e6165c30479a94
calibration/calibration_012.json	c55e17da9b46e28be3d840d9a81ee2316b1e54d94b83a26d38880396050efc6e
calibration/calibration_logs/log_calibration_001.csv	bb12243a53865ead7334755fbaab6e4024df5e776b4e8953746678e1e654ce4c
calibration/calibration_logs/log_calibration_002.csv	090a5efdda145e1f9b6b68b9b71aee65f93ecd65e6c4ff40418086e33d521981
calibration/calibration_logs/log_calibration_003.csv	29ce82637604bd43e538a7cc616bc9fd0706c07d7ef507f528d8b2075344538d
calibration/calibration_logs/log_calibration_004.csv	a2ff46535d906e060272f411c216634b306b1cdff51cd750f8c16599d4f36ca0
calibration/calibration_logs/log_calibration_005.csv	e1937182fd6f8aa3db5aaf84975ab5414c53f02a2bdd7acf18fc34de2f4b34ce
calibration/calibration_logs/log_calibration_006.csv	b953adde5b2ce1b250a808c8eb439bce38c78b06fa07c5a3e6a0bfa8bb95ee66
calibration/calibration_logs/log_calibration_007.csv	88a3468ab3c896b02f13efaef79a3f40e959f649bb798493c4ac41a58692f80b
calibration/calibration_logs/log_calibration_008.csv	3fa3da9842a22719c3e073f8afad990290a244eba297c0e8c701c2a07aec95ec
calibration/calibration_logs/log_calibration_009.csv	b3a00a1bcc6a7608e97c381cbc7c8d96048ac9fe6f328824e81158877a0f4f9b
calibration/calibration_logs/log_calibration_010.csv	d54e979d322ac227b9338d19fecec9718e0b5a406961e28a066387eea109b1ed
calibration/calibration_logs/log_calibration_011.csv	b138fc965b2dbccd5cc44dd3c0004d861551491758c07582a784128e6b2cbba9
calibration/calibration_logs/log_calibration_012.csv	83f2b615a8a63f110794db332b5ff546b068deae0e024362906fbe234ce6bda0
data/frames_preprocessed.jsonl	0b8976a4630bc798cab3140240996080beecf7c6e84d1a5cfe82a884ca38af68
data/frames_raw.jsonl	ec5f0aced5ef84671ea3cbc004286cf334cae8c8b4868b93f88445a9931ee65c
data/labeled.jsonl	b2db598a951483af630af11ed0eb7bce2610527ca05ae8d2b1307c07d0675629
data/labels.csv	e69cfcbee93e9f99a2f333bb991c9dadc6e93df50b677b95c1c8398f227e35e5
data/labels.jsonl	8db7499f684e72b80e8c1089d449bc704c241f4a3b75437939df6ac9868ae7c8

---

Column inventory of the population data collection (12 drivers × 2 sessions, 2026-08), computed over every row of data/study1_data/ on 2026-09-13. Types are the declared types — written into the dataset card as dataset_info.features so load_dataset casts instead of inferring. Where the source file is mixed the JSON types are given in parentheses; the release normalises them (int 0 → false, int → float) because Arrow inference fails on these files (hr_repair_method null→string) or is order-dependent (is_night/is_junction bool/int). The trainers are indifferent to the normalisation — see "Verification" below.

Files in the release:

HF config	Source file	Published as	Rows	Cols	Unit of one row
frames_raw	raw_data.jsonl	data/frames_raw.jsonl	380,990	63	one DataCollector tick (~20 Hz), HR as filtered live
frames_preprocessed	preprocessed_data.jsonl	data/frames_preprocessed.jsonl	380,990	65	same frames, HR rebuilt offline by heart_rate_preprocessing.py (+2 provenance cols)
labels	user_loa_labels.csv	data/labels.jsonl and data/labels.csv	1,446	22	one driver prompt (two per 20 s window)
labeled	labeled_data.jsonl	data/labeled.jsonl	508,282	74	one (frame, label) pair — the training file; frames are duplicated once per label, see §C
—	calibration_data_study/	calibration/calibration_<pid>.json, calibration/calibration_logs/log_calibration_<pid>.csv	12 + 12 files	—	per-driver 60 s calibration baseline and per-tick log — inputs of the HR repair

The release is built for full reproducibility, which fixes the formats: files are published in exactly the shapes the pipeline reads (JSONL frames, CSV labels), so heart_rate_preprocessing.py, build_loa_dataset.py and every trainer run on them unchanged. Two consequences:

• The Hub applies a single file format to all configs of one repo, so the labels config is data/labels.jsonl; data/labels.csv is the identical table in the CSV form build_loa_dataset.py takes (not a config; the build checks the two agree row for row).
• calibration/ is included because frames_raw alone cannot regenerate frames_preprocessed: the HR repair needs each driver's calibration baseline and per-tick calibration log.

scripts/upload_study1_hf.py stages the release, refuses to upload if the row/column counts differ from this table, and with --verify-chain re-runs the pipeline on the staged inputs — the results are recorded in the card:

Verification (2026-09-13, code 14a3be2). heart_rate_preprocessing.py on the released frames_raw + calibration/ reproduces the released frames_preprocessed on all 380,990 rows and every published column; build_loa_dataset.py on that output + labels.csv reproduces the released labeled on all 508,282 rows; train_XLSTM.normalize_row applied to the original labeled_data.jsonl and to the released labeled.jsonl is identical on every row, so a model trained on the release is the model trained on the original. Each config also loads through datasets with the declared types and participantid keeps its zero padding.

The two frame files are byte-identical in every column outside the rPPG block (0 differing cells over 380,990 rows); only heart_rate, hr_delta and the hr_repair_* columns differ.

---

A. Frame files — frames_raw (63 cols) / frames_preprocessed (65 cols)

Columns marked P exist only in frames_preprocessed.

A.1 Identity / session context

Column	Type	Null	Notes
timestamp	string	0 %	Wall-clock HH:MM:SS.mmm, local time, no date. Join to labels via session_id + window_start_timestamp/window_end_timestamp.
session_id	string	0 %	UUID; 24 distinct
participantid	string	0 %	001–012
traffic_seed	int	0 %	CARLA traffic-scenario seed (TRAFFIC_SEED_PLAN in start_experiment.py)
environment	string	0 %	constant city
secondary_task	string	0 %	constant none
modeltype	string	0 %	constant combined (run config)
state_model	string	0 %	constant xlstm (run config)
w_fcd	float	0 %	constant 0.7 (run config)

A.2 Face / driver state (MediaPipe FaceLandmarker, EmotiEffLib, YOLO26)

Column	Type	Null	Notes
face_present	bool	0 %	landmarker found a face
eye_ar	float	0 %	eye aspect ratio (model feature ear)
mar	float	0 %	mouth aspect ratio
gaze_score	float	0 %	z-scored against the 180 s calibration baseline
gaze_score_raw	float	0 %	uncalibrated gaze score
gaze_distracted	bool	0 %	gaze_score above calibrated threshold (mean + 2.5·std)
blink_rate	float	0 %	normalized (Poisson) against calibration mean
blink_rate_raw	float	0 %	blinks · min⁻¹
yawn_rate	float	0 %	normalized
yawn_rate_raw	float	0 %	yawns · min⁻¹
perclos	float	0 %	z-scored
perclos_raw	float	0 %	fraction of time eyes closed
drowsiness_alert	bool	0 %	PERCLOS + MAR rule
emotion	string	0.2 %	one of angry, disgust, fear, happy, sad, surprise, neutral; null = no reading (no face / classifier failure)
emotion_prob	float	0.2 %	confidence of emotion; null iff emotion is null
lab	list<string>	0 %	YOLO26 distraction classes present: subset of face, phone, drink; may be empty
facebox_misses	int	0 %	face-box worker detector misses (cumulative)
facebox_consec_misses	int	0 %	consecutive misses; box goes stale at 4 s

A.3 rPPG heart rate / respiration (MMRPhys, SCAMPS LEF 72×72)

Column	Type	Null	Notes
heart_rate	float	1.6 %	bpm. frames_raw: live-filtered reading. frames_preprocessed: offline-repaired — differs on 80,156 frames (21 %)
heart_rate_raw	float	1.6 %	unfiltered estimator output
hr_delta	float	1.6 %	heart_rate standardized against the per-driver baseline (median / SD of cleaned calibration readings, floor 5 bpm); recomputed in frames_preprocessed
hr_rejected	bool	1.6 %	live 2f-harmonic filter rejected heart_rate_raw
respiratory_rate	float	1.6 %	breaths · min⁻¹. Not a model input — RGB respiration from the synthetic-trained checkpoint was judged noise
rr_delta	float	1.6 %	standardized RR (median / MAD baseline)
rppg_gaps	int	0 %	look-away discontinuities > 1 s that spliced the model window
rppg_dropped	int	0 %	frames lost to a full rPPG queue
rppg_suppressed	int	0 %	readings flagged as gap-contaminated (never discarded in study 1)
rppg_harmonic_rejects	int	0 %	probable 2f rejections
P hr_repaired	bool	0 %	heart_rate was rewritten offline
P hr_repair_method	string	79 %	folded (45,023 — 2f harmonic halved), interpolated (33,597 — outlier replaced), carried (1,536); null when not repaired

A.4 Vehicle / world (CARLA 0.10, via the vehicle-state bridge)

Column	Type	Null	Notes
speed_kmh	float	0 %	ego speed
speed_limit_kmh	float (int/float)	0 %	0 on the first 20 frames of each session (bridge not yet connected), else 30.0
speed_ratio_max	float	0 %	speed_kmh / 150
speed_ratio_limit	float (int/float)	0 %	speed_kmh / speed_limit_kmh; -1 when the limit is unknown
throttle	float	0 %	[0, 1]
brake	float (int/float)	0 %	[0, 1]
steer	float (int/float)	0 %	[−1, 1]
acceleration	float	0 %	magnitude, m · s⁻²
gear	int	0 %
reverse	bool	0 %
hand_brake	bool	0 %	constant False
is_junction	bool (bool/int)	0 %	ego waypoint is in a junction; int 0 only on the first 20 frames of each session
is_night	bool (bool/int)	0 %	constant False (sun above horizon in every session)
traffic_light_state	string	0.1 %	Red / Yellow / Green
lead_distance_m	float (int/float)	0 %	distance to lead vehicle in ego lane, scaled /100 (so ≈ [0, 1]); -1 = no lead vehicle within 100 m (73 % of frames). Sentinel, not zero — 0 would mean contact
headway_s	float	79 %	time headway, s; null = no lead vehicle, or ego below walking pace
precipitation	float (int/float)	0 %	constant 0 (no weather in CARLA 0.10)
fog_density	float	0 %	constant 0.0
headlight	bool	0 %	constant False
fog_light	bool	0 %	constant False
left_indicator	bool	0 %	constant False
right_indicator	bool	0 %	constant False

A.5 Pipeline timing

Present on most frames, absent on a few early ticks per session (hence two key sets in the JSONL).

Column	Type	Notes
frame_dt_ms	float	inter-frame gap
collect_ms	float	perception loop time for this tick
fps_inst	float	instantaneous achieved rate
fps_avg	float	running mean achieved rate — use to identify participant 001's ~4 Hz warm-up (session 77b516f6, windows 1–12)

---

B. labels — user_loa_labels.csv (1,446 rows, 22 cols)

One row per prompt. Each 20 s window carries two prompts (--random-function), so 723 windows → 1,446 rows.

Column	Type	Notes
session_id	string	join key to frames
participantid	string	001–012
window_idx	int	1-based window index within the session
prompt_in_window	int	1 or 2
window_start_ms	int	window start, CARLA sim time
window_end_ms	int	window end, CARLA sim time (= start + 20,000)
window_start_timestamp	string (ISO 8601)	wall-clock window start — join to frame timestamp
window_end_timestamp	string (ISO 8601)	wall-clock window end
selection_timestamp	string (ISO 8601)	when the driver submitted the answer
selection_frame	int	CARLA frame at submission
selection_sim_time	float	CARLA sim time at submission, s
selection_speed_kmh	float	ego speed at submission
functionname	string	the prompted task, one of five: Provide traffic news (307), Respond to a text message (296), Respond to a phone call (292), Change song (276), Provide weather update (275)
user_selected_loa	string	ground truth. Single LoA 0–4 (1,348 rows), or a ;-joined set of acceptable LoAs (98 rows, 6.8 %): 1;2 ×40, 0;1 ×30, 2;3 ×17, 3;4 ×5, 0;1;2;3 ×4, 1;2;3 ×1, and one non-contiguous 0;3
ambient_gain	float	ambient-audio config (constant 0.35)
ambient_seed	int	ambient-audio config (constant 0)
ambient_source	string	ambient-audio config (constant)
environment	string	constant city
secondary_task	string	constant none
modeltype	string	constant combined
state_model	string	constant xlstm
w_fcd	float	constant 0.7

---

C. labeled — labeled_data.jsonl (508,282 rows, 74 cols)

This is the file the xLSTM population model and every per-driver adaptation were trained on. It is fully derived from frames_preprocessed and labels by scripts/build_loa_dataset.py; it is included so the training input is available as-is, without requiring the join to be reproduced.

C.1 How it was generated

1. Each row of labels defines a 20 s window [window_start_timestamp, window_end_timestamp] within one session_id.
2. Every frame of frames_preprocessed whose session_id matches and whose wall-clock timestamp falls inside that window is attached to the label. Frames outside every window (calibration, the gaps between windows) are dropped: 254,141 of 380,990 frames are labelled.
3. The driver's user_selected_loa becomes the target — never the system's own LoA (which is null here anyway, and would be circular in a served session).
4. functionname and FCD are taken from the label row, not the frame. The collector stamps the CLI default (Adjust seat positioning) on every frame; the label records the task the driver was actually asked about, and its FCD vector is looked up in fcd_config.py.

python scripts/build_loa_dataset.py \
    --raw data/preprocessed_data.jsonl \
    --labels data/user_loa_labels.csv \
    --out-jsonl data/labeled_data.jsonl \
    --out-fcd data/processed_data/fcd_out.csv

C.2 Why rows are duplicated

Under --random-function the drive UI asks two prompts per 20 s window, about two different tasks, and the driver answers each separately. Both label rows share the same window bounds, so they select the same frames. Each frame is therefore emitted once per label: identical driver-state and vehicle features, different functionname, FCD, user_loa and segment_id. Taking only the first match would discard half the labels.

Consequences for anyone using the file:

• The unit of analysis is segment_id, not the row. 1,446 segments = 1,446 labels; 508,282 rows = 254,141 distinct frames × 2. Row counts double-count frames; (session_id, timestamp) identifies a physical frame.
• Frames per segment: min 80, median 381, max 396 (the 80-frame segments are participant 001's ~4 Hz warm-up windows, see §A.5).
• A segment carries one label; the frames within it are one time series. Train/validation/test splits must be made at the segment (or window) level — splitting by row leaks a window's frames across splits.
• The two segments of one window are not independent samples of driver state: they differ only in the task asked about. Treat them as such in any analysis of the state features.

C.3 Columns

All 65 frames_preprocessed columns (§A) are passed through verbatim, plus two columns taken from the label row (the frame files do not carry them — see §D for why the collector's own functionname/FCD were dropped):

Column	Type	Notes
functionname	string	the prompted task (five values, as in labels.functionname)
FCD	struct<12 × int>	the task's FCD vector, keys Safety Risk, Increased Safety, Relevance, Magicality, Privacy, Trust, Time Consumption, Repetitiveness, Situational Context, Social Risk, Urgency, Complexity, values 1–5; identical for all rows of a segment

plus seven columns added by the builder:

Column	Type	Notes
segment_id	string	<session_id>|winNNNpM (NNN = window index, M = prompt 1/2) — the unit of one label. Count these, not rows.
user_loa	string	copy of labels.user_selected_loa for this segment (;-joined sets preserved)
Level_1 … Level_5	int	multi-hot of user_loa; Level_k = 1 iff LoA k−1 is in the set

---

D. Columns excluded from the release

Present in the on-disk files, deliberately not uploaded.

File	Column(s)	Why
frames (both)	functionname	Constant Adjust seat positioning — the CLI default ProVoice was started with, not the task the driver was prompted about. Non-null and plausible-looking, so more misleading than an empty column. The prompted task is labels.functionname (and labeled.functionname, which is taken from the label row).
frames (both)	LoA, FCD	100 % null — no model was served in study 1. The collector emits them because in a served session they hold the decision in force at that frame.
frames_preprocessed, labeled	hr_repair_reason	Free-text audit string for each HR repair (the rule that fired, e.g. 2f harmonic of session 52). Redundant with hr_repaired + hr_repair_method for any downstream use; kept only in the local files.
labeled	LoA	100 % null, passed through from the frames (see above). FCD is kept in labeled — the builder overwrites it with the prompted task's vector (§C.3).
labels	emotion, system_action, system_level, system_loa, system_message, system_probs, system_profile, system_fallback, system_fallback_reason, system_fcd	100 % empty — no system prediction was shown to the driver in study 1.

Under calibration/, only the 12 study participants' calibration_<pid>.json and calibration_logs/log_calibration_<pid>.csv are released: the calibration_<pid>_preprocessed.json files are outputs of heart_rate_preprocessing.py (the reproduction chain regenerates them), and participant 998 is a rig test, not part of the study.

Verification (performed at build time by scripts/upload_study1_hf.py)

• frames_raw: 380,990 rows × 63 columns, as documented
• frames_preprocessed: 380,990 rows × 65 columns, as documented
• labels: 1,446 rows × 22 columns, as documented
• labeled: 508,282 rows × 74 columns, as documented
• data/labels.csv (pipeline input) and data/labels.jsonl (Hub config) hold identical rows
• heart_rate_preprocessing.py run on the released frames_raw + calibration/ reproduces the released frames_preprocessed on all 380,990 rows and every published column
• build_loa_dataset.py run on that regenerated file + the released labels reproduces the released labeled on all 508,282 rows and every published column
• train_XLSTM.normalize_row applied to every row of the original labeled_data.jsonl and of the released labeled.jsonl gives identical output on all 508,282 rows -- the dropped columns and the bool/int normalisation are invisible to the trainers
• load_dataset(..., "frames_raw") yields 380,990 rows with the declared types; participantid keeps its zero padding
• load_dataset(..., "frames_preprocessed") yields 380,990 rows with the declared types; participantid keeps its zero padding
• load_dataset(..., "labels") yields 1,446 rows with the declared types; participantid keeps its zero padding
• load_dataset(..., "labeled") yields 508,282 rows with the declared types; participantid keeps its zero padding