Datasets:
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- sports-science
- athlete-monitoring
- load-management
- synthetic
- healthcare
size_categories:
- 10K<n<100K
LoadBrief-50K
LoadBrief-50K is a synthetic dataset of 50,000 athlete-monitoring narratives paired with structured load-management briefs, grounded in published sports-science criteria. Each example maps a plain-language description of an athlete's recent training load, heart-rate variability (HRV), and wellness data to a structured brief that declares a risk level, classifies overreaching state, explains how the fatigue signals interact, recommends training modifications, and flags when medical review is warranted — conditioned on one of three audiences (athlete, coach, or sports scientist).
It was created to train LoadBrief, a fine-tuned Llama 3 8B model for load-management brief generation. No comparable public resource exists: no dataset pairs athlete monitoring data with written interpretive briefs.
Why Synthetic
Real athlete-monitoring data is scarce, proprietary, and privacy-constrained, and the nearest public resource (PMData) provides only raw sensor logs with no natural-language targets. LoadBrief-50K is generated by a rule-based simulator rather than by prompting a language model — no clinical content is invented by an LLM at any point. The expert knowledge needed to label a case correctly already exists in the sports-science literature as explicit, quantitative decision rules; the simulator encodes those rules once and applies them mechanically.
How It Was Generated
The simulator encodes four bodies of published sports science as quantitative thresholds:
- Gabbett (2016) — acute-to-chronic workload ratio (ACWR) zones
- Plews et al. (2013) — HRV suppression cutoffs
- Meeusen et al. (2013) — functional / non-functional overreaching and overtraining-syndrome criteria
- Hooper & Mackinnon (1995) — wellness norms (sleep, fatigue, soreness, mood, stress)
Generation runs backward from the label: each scenario's ground-truth classification is fixed first, then a realistic 28-day time series is sampled whose statistics fall in the ranges the published thresholds associate with that label, and the same rules independently confirm the label holds. Series are sampled with day-to-day noise, athlete-specific variability, and missing data at lower monitoring tiers, so the narratives resemble real records rather than idealized templates. The narrative input and brief output are realized in natural language from templated, paraphrased fragments conditioned on the verified label, with the audience field controlling vocabulary and length. Every example passes an automated quality filter (schema completeness, narrative–label consistency, recommendation specificity, language naturalness); examples below threshold are rejected and regenerated.
This means each example's label is correct by construction with respect to the published criteria — the data and its label are produced from the same rules, so they cannot disagree. (See Limitations for one exception in override scenarios.)
Dataset Structure
Splits
| Split | Examples |
|---|---|
| train | 40,000 |
| validation | 5,000 |
| test | 5,000 |
| total | 50,000 |
An 80/10/10 split generated with random seed 42 for reproducibility. Splits are disjoint.
Fields
| Field | Type | Description |
|---|---|---|
text |
string | The full formatted training text (narrative + audience + brief), ready for causal-LM fine-tuning |
input_narrative |
string | The monitoring narrative given as input |
audience |
string | Target register: athlete, coach, or sports_scientist |
output |
string | The structured load-management brief |
risk_level |
string | Ground-truth overall risk (e.g. low, moderate, high, critical, and qualified variants) |
overreaching_classification |
string | Ground-truth overreaching state (normal_adaptation, functional_overreaching, non_functional_overreaching, overtraining_syndrome, undertraining) |
complexity_tier |
int | 1 = clear-cut signals, 2 = moderate, 3 = conflicting signals |
sport |
string | One of 66 sports (e.g. swimming_distance, basketball, gaelic_football) |
scenario_type |
string | One of 18 scenario types (e.g. taper, acwr_spike, overtraining_syndrome, heat_acclimatization) |
quality_score |
float | Automated quality-filter score for the example |
Composition
The corpus spans 66 sports, 18 scenario types, and three complexity tiers, with each underlying scenario rendered into all three audience registers. Tiers are weighted toward the harder conflicting-signal cases (roughly a third of the corpus) so models see the full range from routine to genuinely ambiguous situations.
Usage
from datasets import load_dataset
ds = load_dataset("tyhob/loadbrief-50k")
print(ds["train"][0]["input_narrative"])
print(ds["train"][0]["output"])
# Filter to a scenario type or audience
conflicting = ds["train"].filter(lambda x: x["complexity_tier"] == 3)
coach_briefs = ds["train"].filter(lambda x: x["audience"] == "coach")
Intended Uses
LoadBrief-50K is intended for training and evaluating models that translate athlete-monitoring data into structured, evidence-based interpretations, and for research on audience-conditioned generation grounded in explicit rules. It is a research and educational resource, not a medical dataset: it teaches the consistent application of established criteria, not the prediction of real-world injury outcomes.
Limitations
- Rule-grounded, not outcome-validated. Labels are correct with respect to the encoded published criteria, not validated against real athlete injury outcomes. Validation against ratings from practicing sports scientists is future work.
- Override-scenario inconsistency. In a subset of scenarios where the intended classification overrides the raw signals (e.g. environmental suppression in heat acclimatization, developmental variation in youth growth spurt, low-load overtraining syndrome), the brief prose was driven by the sampled signals while the label reflected scenario intent — producing briefs whose body text can contradict their own label. This affects primarily these override scenario types; regeneration with consistency-checked templates is identified as future work.
- Synthetic distribution. The data reflects the encoded rules and templates; it inherits any gaps in the published criteria and does not capture the full messiness of real-world monitoring.
Citation
If you use this dataset, please cite the underlying sports-science sources (Gabbett 2016, Plews et al. 2013, Meeusen et al. 2013, Hooper & Mackinnon 1995) and link back to this dataset and the LoadBrief model.
License
Released under CC BY 4.0. The dataset is fully synthetic and contains no real athlete data.