LoadBrief: Athlete Load-Management Brief Generation

LoadBrief is a Llama 3 8B Instruct model fine-tuned to convert a plain-language athlete-monitoring narrative into a structured, audience-conditioned load-management brief. Given a description of an athlete's recent training load, heart-rate variability (HRV), and wellness scores, the model produces a brief that declares an overall risk level, classifies overreaching state, explains how the fatigue signals interact, gives specific training-modification recommendations, and flags when medical review is warranted — written for an athlete, coach, or sports scientist.

Introduction

Wearable technology has made physiological monitoring available to athletes at every level, but data collection has far outpaced interpretation. Translating an elevated acute-to-chronic workload ratio, a multi-day drop in HRV, and several poor wellness scores into a sound training decision requires expert judgment that most athletes and coaches cannot access. General-purpose language models struggle with this task: they apply generic rather than sport-specific thresholds, reason inconsistently when monitoring signals conflict, and cannot reliably hold an audience register through prompting alone.

LoadBrief addresses this by fine-tuning a specialized model on a purpose-built synthetic dataset whose labels are grounded in published sports-science criteria. The model was trained in two stages — supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) — using LoRA. On a held-out test split it recovers the correct risk classification about three-quarters of the time (composite reward 0.74 against a reference ceiling of 0.729; 97% of risk predictions within one class of ground truth), a dramatic improvement over the untuned base model (reward 0.39, 0% exact risk accuracy). A key empirical finding is that SFT alone saturates the rule-derived reward — GRPO adds no measurable benefit — and that the specialization comes at a measurable cost to general benchmark performance, discussed under Limitations.

Data

The model is trained on LoadBrief-50K, a synthetic dataset of 50,000 monitoring-narrative → brief pairs. No public dataset pairs athlete monitoring data with written interpretive briefs, so the data is generated by a rule-based simulator rather than by prompting a language model — ensuring no clinical content is invented. The simulator encodes four bodies of published sports science as quantitative thresholds: Gabbett (2016) for acute-to-chronic workload ratio zones, Plews et al. (2013) for HRV suppression, Meeusen et al. (2013) for overreaching classification, and Hooper & Mackinnon (1995) for wellness norms. Generation runs backward from the label: each scenario's ground-truth classification is fixed first, a realistic 28-day time series is sampled to match, and the same rules independently confirm the label. Each scenario is rendered into all three audience registers, so one simulated athlete yields multiple training pairs sharing ground-truth labels but differing in tone and technical depth.

The corpus was partitioned into an 80/10/10 train/validation/test split (40,000 / 5,000 / 5,000 examples) generated with random seed 42 for reproducibility. The 5,000-example test split is disjoint from training and is used for all in-domain evaluation reported below. Scenarios span three complexity tiers (clear-cut, moderate, and conflicting-signal cases) across 66 sports and 18 scenario types.

Methodology

The base model is Llama 3 8B Instruct, fine-tuned with LoRA so the full pipeline runs on a single commodity GPU. Training proceeds in two stages. An SFT stage first teaches the structured brief format and the three audience registers from the labeled input–output pairs. A GRPO stage then optimizes a five-component composite reward — correct risk-level declaration, recommendation specificity, correct overreaching terminology, presence of an escalation trigger, and clinically useful output length — by sampling several candidate briefs per prompt and reinforcing the higher-scoring ones. GRPO is chosen over PPO because it removes the separate value network (roughly halving memory), and over preference methods such as DPO because the task has a programmatic reward function rather than human preference pairs.

Hyperparameters (for reproducibility). SFT stage: LoRA rank 32, alpha 64, dropout 0.05, applied to all attention and MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj); learning rate 2e-4 with a cosine schedule, 3% warmup, weight decay 0.01; per-device batch size 2 with gradient accumulation 16 (effective batch 32); 1 epoch over 20,000 training examples; max sequence length 512; gradient checkpointing enabled; best checkpoint selected by validation loss. GRPO stage: learning rate 5e-6, KL coefficient (β) 0.1, sampling temperature 0.8, top-p 0.95, generation seed 42, LoRA configuration inherited from the SFT stage. A learning-rate sweep (5e-5, 5e-6, 5e-7) and multiple random seeds confirmed the results are stable across configurations; the released model uses the 5e-6 GRPO configuration.

Evaluation

The model is evaluated on three standard benchmarks plus the held-out LoadBrief test split. MedQA (USMLE) probes domain-adjacent clinical reasoning; MMLU (clinical knowledge, college medicine, and philosophy subjects) measures retained general and medical knowledge, with philosophy as a domain-unrelated control; and the LoadBrief test split measures in-domain performance via composite reward and risk-classification accuracy. Two similar-size instruction-tuned models are included as comparisons: Qwen 2.5 7B Instruct, a strong recent 7B model, and Mistral 7B Instruct v0.3, a widely-used baseline 7B — both represent the general-purpose alternatives a practitioner might otherwise reach for.

Model	MedQA	MMLU clinical	MMLU college med	MMLU philosophy	LoadBrief test (reward / risk acc)
Llama 3 8B Instruct (base)	61.7	77.4	63.6	71.7	0.39 / 0.00
LoadBrief (SFT+GRPO)	55.8	66.0	59.5	59.8	0.74 / 0.72
Qwen 2.5 7B Instruct	61.7	77.4	67.6	73.6	—
Mistral 7B Instruct v0.3	51.9	68.7	59.0	67.2	—

Benchmark values are accuracy percentages. LoadBrief test reports composite reward (reference ceiling 0.729) and exact risk-level accuracy. Comparison models are not trained on the LoadBrief format and, like the untuned base model, score at floor on the in-domain task, so that column is omitted for them.

On the general benchmarks, the comparison models match or exceed LoadBrief — Qwen 2.5 7B is strongest (72.9% MMLU average) while LoadBrief sits lowest (61.8% average), reflecting the general capability traded away during task specialization. However, none of the comparison models produces valid structured briefs: on the in-domain task only the fine-tuned model performs meaningfully (0.74 reward, 0.72 risk accuracy) versus the base model's floor (0.39, 0.00). In short, the specialization that lowers general-benchmark performance is precisely what enables the target task.

Usage and Intended Uses

LoadBrief is a LoRA adapter applied to Llama 3 8B Instruct. Load it with the Transformers and PEFT libraries:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Meta-Llama-3-8B-Instruct"
ADAPTER = "tyhob/loadbrief"  

tokenizer = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

prompt = """### Monitoring Narrative:
Weekly load management brief (4-week monitoring block): Soccer Midfielder, professional.
Acute:chronic workload ratio of 1.71. HRV suppressed 13 ms below baseline for 6
consecutive days. Wellness moderately depressed across sleep, fatigue, and mood.

### Audience: coach

### Load Management Brief:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=400, temperature=0.7)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Intended uses. LoadBrief is intended as a research demonstration and educational tool for translating athlete-monitoring data into structured, evidence-based interpretations across audiences. It is well suited to generating first-draft briefs that a coach or sports scientist reviews. It is not a medical device and should not be used for autonomous clinical or return-to-play decisions. The model applies established sports-science criteria consistently; it does not predict real-world injury outcomes.

Prompt Format

Each prompt contains three parts: the monitoring narrative, an audience specifier (athlete, coach, or sports_scientist), and a trailing header that cues generation.

### Monitoring Narrative:
Weekly load management brief (4-week monitoring block): Subject: Canoe Kayak Sprint
Kayak Single (semi professional), 10-year training background, return from injury phase.
Training load is well-periodized — acute:chronic workload ratio of 1.256 reflects a
healthy balance between fitness and fatigue. Full wellness questionnaire not available
— sleep data from wearable only.

### Audience: athlete

### Load Management Brief:

Expected Output Format

The model returns a structured brief with a risk header, per-signal sections (training load, HRV, wellness), a signal-integration analysis, recommendations, and an escalation trigger. Vocabulary, headers, and length adapt to the audience: athletes receive a YOUR STATUS: header with lay terms ("recovery score," "a few easy days"), while coaches and sports scientists receive a RISK LEVEL: header with clinical terminology and exact values. The following is a sports-scientist brief:

RISK LEVEL: HIGH

TRAINING LOAD:
ACWR of 1.14 is within the optimal training range. Acute load: 106 AU | Chronic load: 94 AU.

PHYSIOLOGICAL READINESS (HRV):
HRV critically suppressed: 23.3 ms (22.4 ms below baseline for 7 consecutive days).

SUBJECTIVE WELLNESS:
Composite status: moderately depressed. Amber flags: sleep, fatigue, soreness, mood.

CLINICAL CLASSIFICATION:
Functional overreaching — short recovery period (3–7 days) will restore performance.

RECOMMENDATIONS:
- If a non-training stressor is identified: maintain load, address the stressor.
- If none: reduce load 20% and reassess in 3 days; if no improvement, seek medical review.

ESCALATION TRIGGERS:
Persistent multi-system suppression, performance decline, or athlete distress.

Limitations

General-capability forgetting. Fine-tuning measurably degraded performance on general and clinical-reasoning benchmarks: MedQA fell 61.7% → 55.8%, and MMLU subjects dropped by 4–12 points (clinical knowledge −11.3, philosophy −11.9). Philosophy — the benchmark least related to the training domain — dropped the most, indicating the degradation is general rather than a targeted reshaping of clinical knowledge. LoadBrief is a specialist and should not be relied upon for general or clinical question-answering outside its task.

Weakness on override scenarios. A stratified LLM-as-judge evaluation found clinical accuracy is weakest on scenarios where the intended classification overrides the raw signals — overtraining syndrome, heat acclimatization, and youth growth spurt score lowest (~2.2–2.6 of 7) versus ~4.0–4.7 for signal-clear cases. This traces to a data-generation limitation: in these scenarios the brief prose was driven by the sampled signals while the label reflected scenario intent, producing internally inconsistent briefs the model learned to reproduce. Regenerating these scenarios with consistency-checked templates is identified as future work.

Rule-grounded, not outcome-validated. The model learns to apply established sports-science criteria consistently; its labels are correct by construction with respect to those criteria, but are not validated against real-world injury outcomes. Validation against ratings from practicing sports scientists is future work.

SFT saturation. The GRPO stage adds no measurable benefit over SFT alone on this task, because the rule-derived reward is already saturated after supervised fine-tuning. Users seeking to reproduce or extend this work can achieve equivalent quality with the SFT stage alone.

Training-data subset. SFT used a 20,000-example subset of the 40,000-example training split, chosen to fit within the memory budget. Training compute scales linearly with example count, so the full split was tractable but not necessary: as noted above, the model reached the reference reward ceiling at this scale, indicating additional in-distribution data would yield diminishing returns. Targeted augmentation of under-performing override scenarios, rather than more uniform data, is the more promising direction for future improvement.

---

Built with Meta Llama 3. Model weights are subject to the Meta Llama 3 license. Training code and the data-generation pipeline are available on GitHub.