tyhob/loadbrief-50k

LoadBrief

A synthetic corpus of athlete load-management monitoring briefs. Each record pairs a free-text monitoring narrative (training load, heart-rate variability, subjective wellness) with structured coaching briefs in three audience registers.

This card documents version v8, the corpus used to train tyhob/loadbrief.

Read the Label Provenance and Known Defects sections before training on this corpus. Both ground-truth labels are per-scenario constants rather than functions of the sampled physiological signals, and 16.1% of records are internally self-contradictory. These properties are documented rather than repaired, so that the corpus remains the corpus of record for the released model. An accompanying paper analyses them in full.

Dataset structure

Field	Description
input_narrative	Free-text monitoring report for a 28-day block
output_athlete	Structured brief, athlete register
output_coach	Structured brief, coach register
output_sports_scientist	Structured brief, sports-scientist register
risk_level	Ground-truth risk label
overreaching_classification	Ground-truth overreaching class
acwr_value, acwr_zone	Acute-to-chronic workload ratio and its zone
complexity_tier	1 (clear-cut), 2 (moderate), 3 (conflicting signals)
conflicting_signals	Whether the scenario declares an intentional signal conflict
scenario_type	One of 19 scenario labels
sport, sport_category, athlete_level, training_phase	Athlete context
data_completeness_level	1–4, which monitoring signals are available
quality_score	Composite generation-quality score (threshold 0.60)

Briefs use a fixed section layout: RISK LEVEL / YOUR STATUS / OVERALL RISK CLASSIFICATION (register-dependent), then TRAINING LOAD, PHYSIOLOGICAL READINESS (HRV), SUBJECTIVE WELLNESS, SIGNAL INTEGRATION, CLINICAL CLASSIFICATION, RECOMMENDATIONS, MONITORING PRIORITIES, ESCALATION TRIGGERS.

risk_level takes values low, low_if_managed, low_with_detraining_flag, moderate, moderate_to_high, high, critical, variable. overreaching_classification takes normal_adaptation, functional_overreaching, non_functional_overreaching, overtraining_syndrome.

Splits: 40,000 train / 5,000 validation / 5,000 test, an 80/10/10 partition with seed 42. 66 sports across 10 categories, 19 scenario types, 11 training phases.

Generation method

Records are produced by a rule-based simulator. Scenario parameters are sampled, a 28-day time series of loads, HRV and wellness is simulated to match the scenario's declared trajectory, derived metrics are computed, and briefs are rendered from the surviving signals. No language model participates, so no clinical content is fabricated. Thresholds follow Gabbett (ACWR), Plews et al. (HRV), Meeusen et al. (overreaching), Hooper & Mackinnon (wellness) and Foster (monotony).

Generation runs backward from the label: the scenario's classification is fixed first and the time series is sampled to match. Everything below follows from that choice.

Label provenance

Both ground-truth labels are per-scenario constants. risk_level and overreaching_classification are written directly from the scenario definition and are not functions of the sampled signals. Across all 40,000 training records, each of the 19 scenarios maps to exactly one risk label and one overreaching class, with zero within-scenario variance.

The practical consequence: classification accuracy on this corpus measures recovery of the generating scenario, not clinical inference. A TF-IDF bag-of-words classifier with logistic regression recovers scenario_type from input_narrative at 0.905 (majority baseline 0.071) and risk_level at 0.950 — within one point of a fine-tuned Llama 3 8B. Report accuracy on this corpus accordingly.

Known defects

1. Two families of self-contradictory records (6,433 records, 16.1%)

The label cannot respond to the sampled signals, so wherever the sampler disagrees with the declaration the brief contradicts itself. This surfaces on two disjoint axes; no record carries both.

Risk axis (2,412 records, 6.0%). The brief reports critically suppressed HRV in PHYSIOLOGICAL READINESS while declaring RISK LEVEL: LOW. SIGNAL INTEGRATION typically acknowledges the discrepancy and advises trusting HRV over the subjective report, and the header declares low risk anyway.

Classification axis (4,021 records, 10.1%). SIGNAL INTEGRATION names one overreaching class while CLINICAL CLASSIFICATION declares another — adjacent paragraphs asserting different classifications.

Scenario	n	Risk axis	Class axis	Rate
acwr_spike	2,824	0	2,416	85.6%
undertraining	2,836	1,207	0	42.6%
preseason_intensification	2,425	0	960	39.6%
post_competition	2,656	1,038	0	39.1%
monotony_problem	2,396	0	327	13.6%
high_acwr_stable_physiology	1,733	0	211	12.2%
normal_progressive	2,767	165	0	6.0%
fixture_congestion	2,506	0	107	4.3%
taper	2,722	2	0	0.1%

The other ten scenarios are clean. Per-scenario contradiction rate correlates with independently judged clinical accuracy at Spearman ρ = −0.67, both for model outputs and for these reference briefs.

To filter the risk axis, drop records whose output_coach contains both "critically suppressed" and "RISK LEVEL: LOW". For the classification axis, compare the overreaching terms appearing in the SIGNAL INTEGRATION section against those in CLINICAL CLASSIFICATION and drop records where they disagree. Dropping all nine affected scenarios removes 16.1% of the corpus.

2. Three consistency guards pass every affected record

The pipeline includes a label-agreement filter, a phrase-bank consistency checker, and a schema validator. All three accept all 6,433 records, each for a different structural reason: the filter operates only on the overreaching axis and ranks undertraining at the same severity as normal_adaptation; the checker exempts briefs containing conflict-acknowledgment language, which these contain; and the validator's cross-field check is one-directional, verifying only that severe labels appear in the text. Passing these checks is not evidence of consistency.

3. CRITICAL is unreachable above ACWR 0.8

All 1,706 records labeled critical are exactly the 1,706 overtraining_syndrome records, and that scenario declares an ACWR target of 0.45–0.65 (the depleted phase). No record above ACWR 0.8 is labeled critical under any physiological condition. A model trained on this corpus will not assign CRITICAL to a high-load presentation.

4. The overreaching label is not readable by string extraction

From v2 onward, conflicting-signal classes carry a rationale paragraph in CLINICAL CLASSIFICATION instead of a class name. Passing the corpus's own ground-truth briefs through a standard extraction metric recovers only 55% of correct answers, and 0 of approximately 5,100 overtraining_syndrome records in any register. non_functional_overreaching recovers at 0.33 and normal_adaptation at 0.49.

If you evaluate overreaching classification on this corpus, score against the overreaching_classification field directly. Do not parse it out of generated text: 0.55 is the ceiling any such metric imposes, not 1.0. The label itself is recoverable from the narrative at 0.971 by a linear classifier, so the information is present.

5. early_overreaching carries an unresolved label–signal divergence

All 1,716 early_overreaching records declare non_functional_overreaching while their sampled ACWR (0.9–1.2) lies in the sweet-spot zone, from which the rule-derived class is normal_adaptation — a two-rank severity gap in every record. These are exempted from the label-agreement filter by their conflicting_signals: true flag rather than repaired. Treat the flag as a declaration of intent, not as evidence the divergence was resolved.

6. The athlete register carries less clinical content

Ground-truth athlete briefs score 0.88 lower on judged clinical accuracy than coach briefs (3.52 vs 4.40 on a 1–7 scale), and are lowest of the three registers in 16 of 19 scenarios. The audience adapter substitutes lay vocabulary ("recovery score" for HRV, "overtraining warning" for non-functional overreaching), which meets its readability targets but removes clinical content rather than restating it.

7. Declared-target reachability

v8 satisfies its own declared ACWR and monotony ranges for 18 of 19 scenarios. recreational_minimal_data declares 0.8–1.6 and produces 0.52–1.20, with 27% of samples in range. Earlier revisions were substantially worse; nine targets were unreachable at v1.

8. Declared tier shares do not match the corpus

The generator builds its scenario pool by repeating the tier-1, tier-2 and tier-3 scenario lists 8, 7 and 5 times. Those multipliers give 40/35/25 only if the three lists are equal in length; they hold 5, 5 and 9 scenarios, so the pool shares are 33.3% / 29.2% / 37.5%. Realized shares in the released corpus are 34.5% / 30.2% / 35.3% (13,805 / 12,063 / 14,132 records), the residual being uneven rejection by the quality filter.

Tier 3 is therefore over-represented relative to the intent recorded in the source, and tier 1 under-represented. The effect on any downstream use is small, but if you weight or stratify by complexity tier, use the realized counts above rather than the nominal 40/35/25.

9. Fully synthetic, not outcome-validated

No real athlete data was used. The corpus has not been validated against clinical practice, real monitoring data, injury outcomes, or expert review. The underlying sports-science constructs — the ACWR in particular — are themselves contested in the literature.

Intended use

Research on structured generation, register control, and synthetic-corpus quality — including as a worked example of how rule-generated data fails. Not suitable for training systems that inform real athlete health or training decisions.

Versioning

This corpus is published unmodified as the corpus of record for sft_checkpoint_v8_final. The defects above are documented rather than repaired so that the dataset and the released model remain consistent. Earlier revisions (v1–v7) are not published.

Citation