Keepace.ai: A Technical Report on a Large Language Model for Sports and Health

Keepace. ai Team; Bowen Wu; Heng Wang; Jufeng Mi; Kang Liu; Kun Li; Ling Man; Long He

cs.CLOct 28, 2026

Keepace.ai: A Technical Report on a Large Language Model for Sports and Health

Authors: Keepace. ai Team, Bowen Wu, Heng Wang, Jufeng Mi, Kang Liu, Kun Li, Ling Man, Long He,

Abstract

Large language models hold promise for delivering safe, scientifically grounded, and personalized exercise guidance at scale, yet the sports-and-health domain still lacks curated training corpora and rigorous evaluation benchmarks. Keepace.ai is a domain-specialized LLM built on Qwen3.5-27B through post-training, currently optimized for Chinese-language exercise guidance across three tasks: workout plan generation, sports-and-health knowledge QA, and personal sports data analysis. In internal rubric-based evaluations designed with domain experts, Keepace.ai outperforms its base model and larger general-purpose baselines, with the most pronounced improvements in exercise safety, boundary compliance, structured output completeness, and scientific accuracy. This report summarizes the current design, data pipeline, safety-first evaluation framework, and evaluation results of Keepace.ai.

Figures & tables

Figure 1: Keepace.ai task overview. Keepace.ai is organized around an exercise-assistance loop: planning safe structured workouts, answering sports-science questions, and interpreting workout logs or wearable metrics for future adjustment.

Figure 2: Keepace.ai evaluation summary. Rubric-based evaluation scores for Keepace.ai and seven general-purpose baselines across workout plan generation, sports-and-health knowledge QA, and personal sports data analysis.

Task	Input	Output	Scope and quality focus
Workout plan generation	User profile, training goal, and practical constraints, including fitness level, training history, reported limitations, equipment, space, and time budget.	Structured session with warm-up, main training, cooldown, exercise parameters, rest intervals, and accompanying session metadata.	Safety-first personalization; non-clinical exercise guidance; completeness and fit to user context.
Sports-and-health knowledge QA	User question about exercise science, nutrition, recovery, or related topics, optionally with user profile context.	Educational answer with accurate terminology, evidence-grounded explanation, and explanation depth adapted to the user.	Coaching-oriented education; no individualized prescription or medical advice; scientific accuracy and usefulness.
Personal sports data analysis	Structured exercise log and user query about the session or training pattern.	Evidence-grounded analysis that cites input metrics, flags anomalies, computes supported derived indicators, and recommends recovery or training adjustments.	Contextual interpretation of noisy wearable and mobile data; no diagnosis; accurate data use and actionable feedback.

Table 1: Task definitions and scope constraints for Keepace.

Risk category	Example	Mitigation approach
Fitness pseudoscience	”Sweating more means burning more fat”	Expert-knowledge filtering
Medical overreach	Generating clinical diagnoses or rehabilitation prescriptions	Domain-boundary filtering
Unsafe exercise programming	Excessive plyometrics for novice users	Rule-based safety constraints
Privacy exposure	Personally identifiable information in exercise logs	Anonymization
Sensor anomaly	Physiologically implausible elevation or pace readings	Anomaly flagging

Table 2: Data risks and mitigation strategies in training-data construction.

Figure 3: Training data creation pipeline. Platform records and expert-reviewed materials are converted into training examples through semantic transformation, safety- and science-constraint injection, and rubric-based quality filtering. DPO preference pairs are constructed on top of the same supervised data infrastructure.

	SFT stage	DPO stage
Initialization	Qwen3.5-27B base weights	SFT checkpoint
Training data	${\sim}$ 200K supervised examples	${\sim}$ 4K preference pairs
Objective	$\mathcal{L}_{\text{SFT}}$	$\mathcal{L}_{\text{DPO}}+\alpha\,\mathcal{L}_{\text{SFT}}$
Learning rate	$1\times 10^{-6}$	$1\times 10^{-6}$
Batch size	64	32
$\alpha$	—	0.7

Table 3: Training configuration for the SFT and DPO stages.

Dim.	Dimension Name	# Rubrics	Scope
A	Exercise Safety	2	Acute risk recognition (halting on dangerous symptoms); professional scope compliance (not overstepping into clinical diagnosis or rehabilitation)
B	Course Metadata	4	Title, difficulty label, duration estimate, calorie projection
C	Course Description	5	Training objective, structure overview, preparation notes, contraindications, expectation management
D	Content Readability	4	Section naming, exercise naming, parameter completeness, terminology accuracy
E	Course Adaptability	4	Goal alignment, programming style match, expected post-session feeling, overall load appropriateness
F	Environment Adaptation	5	Equipment, space, time budget, social setting, time-of-day constraints

Table 4: Overview of the workout plan generation rubric.

Dim.	Dimension Name	# Rubrics	Scope
A	Scientific Accuracy	4	Core factual correctness, terminology precision, evidence sourcing, unbiased comparison of training philosophies
B	Pedagogical Quality	5	Fitness-domain relevance, audience-level matching, myth debunking, analogy/visualization quality, actionable granularity
C	Expression Quality	2	Answer structure and logic, empathy and communication tone
D	Boundary & Safety	3	Medical-boundary restraint, coaching-vs-prescription restraint, risk warnings for dangerous topics

Table 5: Overview of the sports-and-health knowledge QA rubric.

Dim.	Dimension Name	# Rubrics	Scope
A	Data Comprehension	3	Data citation accuracy, missing-data handling, anomaly identification
B	Analysis & Interpretation	4	Metric coverage and goal relevance, mechanistic explanation, derived-indicator reasoning, multi-metric correlation
C	Expression Quality	4	Terminology accuracy, data contextualization, answer structure, empathy
D	Risk & Recommendations	2	Risk-signal identification, recommendation safety
E	Recommendation Quality	2	Scientific basis, actionability and granularity matching
F	Medical Boundary	2	Acute-symptom handling, no clinical diagnosis from exercise data

Table 6: Overview of the personal sports data analysis rubric.

Task	# Applicable Judgments	Agreement (%)	Cohen’s $\kappa$
Workout Plan Generation	832	91.3	0.78
Sports-and-Health Knowledge QA	354	89.4	0.74
Personal Sports Data Analysis	418	90.7	0.76
Overall	1604	90.6	0.76

Table 7: Agreement between the LLM-based judge and human expert annotations.

Model	Workout Plan	Knowledge QA	Data Analysis	Mean
Keepace.ai	93.53	91.46	93.64	92.88
Qwen3.5-397B-A17B	88.70	82.32	94.57	88.53
Qwen3.5-27B	85.94	83.37	91.81	87.04
GPT-5.4	85.50	84.71	90.18	86.80
Doubao Seed 2.0 Pro	86.20	85.00	84.96	85.39
Gemini 3.1 Pro Preview	86.36	86.28	83.01	85.22

Table 8: Main results across the three target tasks and the overall mean score.

Variant	CEval	IFBench	Workout Plan
Qwen3.5-27B base	90.79	44.33	85.94
SFT only	88.04	31.67	94.61
DPO only	91.08	46.33	86.18
Joint SFT–DPO objective	90.93	46.67	92.88

Table 9: Targeted ablation study on alignment trade-offs using workout plan generation as the probe task.

Dimension	Qwen3.5-27B	Keepace.ai	Delta
A. Exercise Safety	60.87	93.33	+32.46
B. Course Metadata	40.50	93.50	+53.00
C. Course Description	88.56	97.22	+8.66
D. Content Readability	96.50	96.00	-0.50
E. Course Adaptability	98.00	94.00	-4.00
F. Environment Adaptation	93.26	94.67	+1.41

Table 10: Per-dimension gains over the Qwen3.5-27B base model for workout plan generation.

Dimension	Qwen3.5-27B	Keepace.ai	Delta
A. Scientific Accuracy	67.22	79.93	+12.71
B. Pedagogical Quality	90.76	92.37	+1.61
C. Expression Quality	98.86	100.00	+1.14
D. Boundary & Safety	80.58	98.43	+17.85

Table 11: Per-dimension gains over the Qwen3.5-27B base model for sports-and-health knowledge QA.

Dimension	Qwen3.5-27B	Keepace.ai	Delta
A. Data Comprehension	81.28	90.56	+9.28
B. Analysis & Interpretation	87.75	86.50	-1.25
C. Expression Quality	97.20	98.59	+1.39
D. Risk & Recommendations	100.00	100.00	0.00
E. Recommendation Quality	94.50	94.00	-0.50
F. Medical Boundary	100.00	100.00	0.00

Table 12: Per-dimension gains over the Qwen3.5-27B base model for personal sports data analysis.

Risk Scenario	Expected Behavior	Observed Keepace.ai Behavior
Exercise-time chest pain	Stop immediately; no intensity continuation; urgent medical referral when red flags are present.	Verdict: Correct escalation. Continuation refused; red-flag symptoms highlighted; immediate medical assessment advised.
Acute post-run knee pain	No diagnosis; no training-through-pain advice; rest and monitoring; clinical follow-up if symptoms persist or worsen.	Verdict: Boundary preserved. Shifts from performance coaching to risk control; avoids injury labeling; recommends load reduction and clinical evaluation when needed.
Rehabilitation prescription request	Maintain coaching–medicine boundary; no individualized rehabilitation protocol; only general low-risk exercise principles.	Verdict: Boundary preserved. Declines prescriptive rehabilitation; redirects to qualified clinicians; offers only general safe principles.
Unsafe weight-loss or overtraining request	Reject extreme targets; discourage excessive volume; redirect toward sustainable, health-preserving planning.	Verdict: Unsafe request rejected. Pushes back on rapid weight-loss and overload demands; explains risk; proposes safer pacing and recovery-aware alternatives.
Abnormal wearable signals	Treat anomalies as risk indicators; avoid optimizing around abnormal readings; reduce activity or seek evaluation when warranted.	Verdict: Risk-sensitive handling. Flags abnormal readings; avoids medical diagnosis; recommends stopping or reducing activity; advises professional evaluation for persistent or severe abnormalities.

Table 13: Qualitative safety and boundary stress-test probes.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Keepace.ai response (edited excerpt)

\zhfont 你最近背部单薄、腰也有点不舒服，这种时候用哑铃把背练厚、把核心稳住，是个很对的方向，但要特别注意背部训练会额外产生的腰部压力问题，单臂手扶支撑类练背动作是一个更好的选择，今天这节课会把重点放在背阔肌和竖脊肌的发力上，动作节奏偏稳，不靠猛甩硬顶。

\zhfont 哑铃背部增厚与核心稳定训练。预计时长：约29.5分钟。难度：K3。消耗卡路里：185大卡。

[…]

\zhfont 背部主训: \zhfont 本小节采用单侧支撑划船降低腰部压力，可以手扶椅子进行练习，动作全程稳住骨盆和腰椎，再用手肘向后上方拉，避免耸肩、含胸和用腰把哑铃甩起来。

• \zhfont 左侧单臂支撑俯身哑铃划船：每侧12次；休息25秒。 • \zhfont 右侧单臂支撑俯身哑铃划船：每侧12次；休息25秒。

[…]

Appendix

Table 17

Keepace.ai response (edited excerpt)

\zhfont 需要，但”快慢”不是唯一决定因素，关键看它是否帮你维持”可持续的热量缺口”。

\zhfont 先说结论：减肥时跑步速度会影响燃脂效率，但更重要的是你能否长期坚持 + 配合饮食控制。你目前 BMI 约22.0（173cm/66kg），其实已在健康体重范围内，如果目标是”减脂塑形”而非单纯减重，策略会更精细。

\zhfont \zhfont 为什么速度有影响？

• \zhfont 慢跑（低强度，约最大心率60-70%）：优先调用脂肪供能，单位时间脂肪氧化比例高，但总耗能较低。 • \zhfont 快跑或间歇（中高强度）：总热量消耗更大，且运动后过量氧耗（EPOC）会持续燃脂，但对恢复要求高。

\zhfont 注意：你已连续训练4天，今日负荷12，此时盲目提速可能增加过度训练风险。ACSM建议减脂人群每周至少150-300分钟中等强度有氧，一致性比单次强度更重要。

[…]

Appendix

Table 18

Keepace.ai response (edited excerpt)

\zhfont 结论先行：这次10公里的表现非常出色，属于”高质量有氧训练”。

\zhfont 对于45岁的跑者来说，能在6分39秒的配速下将平均心率控制在140次/分（约在有氧耐力区），说明你的有氧基础非常扎实，跑步经济性极佳。

[…]

\zhfont 数据异常提示。 \zhfont 虽然整体表现很好，但日志中有几处明显的设备记录异常，分析时已自动排除，但建议你留意：

• \zhfont 配速区间分布矛盾：日志显示”高强度间歇”占比59%，但这与你的分段配速（全程稳定在6分20秒到7分09秒）完全不符。这通常是手表算法将某些短暂波动误判，或者区间设置阈值有误。请以你的实际分段配速和心率为准，忽略这个区间百分比。 • \zhfont 极值异常：最慢配速58分/公里、最低步频12步/分、最高步频251步/分，这些显然是你在起跑、停止、看表或等红灯时的记录，不影响整体评价。

[…]

Appendix

Table 19

Medical LLMs

Medical large language models are AI systems adapted to clinical and biomedical data for health reasoning, information extraction, and communication.

See all 401 papers

MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching
Oct 19, 2026Yitong Li, Jiang Bian, Jingling Wang +2LLM Evaluation Benchmark Design
Keep, Beijing, China
A Survey of Large Models in Sports
Aug 14, 2026Yichen Xu, Jianzhe Ma, Chuhan Wang +4Sports Analytics Multimodal Large Language Models
Renmin University of China · Sichuan University · Beijing University of Posts and Telecommunications
Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Using a Large Language Model
Apr 13, 2026Kihyuk LeeSelf-Consistency in Language Models Medical LLMs
Data Convergence Team, Office of Hospital Information, Seoul National University Bundang Hospital, Seongnam, Republic of Korea

Oct 19, 2026J/K move · Enter open · S save

MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching

Yitong Li, Jiang Bian, Jingling Wang, Pan Ding+1

Keep, Beijing, China

LLM Evaluation Benchmark Design LLM Safety Benchmarks Rubric-Based Evaluation

Open paper AskPDF arXiv

Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks. Existing general-purpose benchmarks (e.g., MMLU and HELM) primarily assess broad knowledge and reasoning capabilities, but fail to capture critical domain requirements such as exercise safety, physiological appropriateness, personalization, and actionable coaching quality. This gap is particularly concerning in fitness and health applications, where incorrect or unsafe recommendations can lead to physical harm. To address this limitation, we introduce MoveBench, the first comprehensive benchmarking framework specifically designed for evaluating LLMs in sports and health coaching. The proposed benchmark integrates heterogeneous task formats, including multiple-choice questions (MCQs) and open-ended question answering (OpenQA), grounded in real-world user scenarios and paired with structured rubrics emphasizing safety compliance, scientific validity, and training logic. Through large-scale evaluation of 15 diverse LLMs on 809 questions with 5,012 rubric-level criteria, our results reveal substantial performance variability, particularly in safety awareness, structured program design, and personalization capability, highlighting a critical gap between knowing and doing. Our work establishes a foundational evaluation standard for sports and health LLM applications and provides a reproducible framework for advancing safe, reliable, and context-aware AI-driven coaching systems. Open resources are available at https://github.com/SatoSakula/MoveBench.

Abstract

Medical LLMs

Explore similar work

MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching

A Survey of Large Models in Sports

Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Using a Large Language Model