lia-prop13/pilates-dataset

---

Project Overview

THE PILATES SESSION is an end-to-end data science project that transforms the classical Pilates mat repertoire into an intelligent, AI-powered lesson-planning engine.

The project follows a structured data science workflow, beginning with the creation of a synthetic dataset - generated through careful prompt engineering on top of a clinical knowledge base - to capture the classical repertoire in a rich, structured form. This is followed by a thorough Exploratory Data Analysis (EDA) to understand the patterns, relationships, and quality of the data, and to refine it into a reliable foundation. Together, these stages build the data backbone of an AI application that generates safe, tailored Pilates lesson plans on demand.

This repository documents that foundation - from the real-world problem that motivated it, through the data generation process, to the insights uncovered during analysis. The project continues in the Space README, where this data comes to life as a live, interactive application.

---

Quick Links

• Dataset Viewer - Explore the full dataset
• Notebook - Data generation & EDA
• Live Application - Try THE PILATES SESSION on Hugging Face Spaces
• Space README - Full application walkthrough

---

Background & Motivation

Background - The Method & Its Founder

Joseph Hubertus Pilates (1883–1967) was a German-born physical trainer who created the exercise system he called Contrology — known today simply as Pilates. A sickly child who battled asthma and rickets, he devoted himself to physical conditioning and became a gymnast, boxer, and diver. During his internment in England as a German national in World War I, he began refining a system of mat exercises, reportedly rigging bed springs to help bedridden patients rebuild strength - the seed of the spring-based apparatus he would later invent.

In 1926 he emigrated to New York and, with his wife Clara, opened a studio that quickly drew dancers and performers, including George Balanchine and Martha Graham. Over his career he developed more than 600 exercises across the mat and his invented apparatus — the Reformer, the Cadillac, and others — all built on a single philosophy: that the body should be trained as an integrated whole, coordinated from a strong center he named the "Powerhouse." His 1945 book, Return to Life Through Contrology, codified the 34 classical mat exercises that remain the canonical repertoire taught to this day.

This classical repertoire - anatomically deliberate, carefully sequenced, and time-tested - is the foundation on which THE PILATES SESSION is built.

The Problem

Joseph Pilates never trademarked his method, and in 2000 a U.S. federal court made that absence official: it ruled that "Pilates" is a generic term - like "yoga" or "aerobics" - that no single entity can own. The decision democratized the method, but it also removed any gatekeeping. Today there is no central authority governing who may teach Pilates, or how. Anyone can open a teacher-training program and certify instructors by their own standards.

The downstream consequences are real:

• Newly certified instructors enter the market - and obtain professional liability insurance - with no guarantee that the training they completed was rigorous or safety-oriented.
• Under constant pressure to keep classes fresh, some improvise exercises with little regard for their biomechanical consequences, drifting away from the anatomically-grounded classical repertoire.
• Clients with pathologies - disc herniation, osteoporosis, pregnancy, joint conditions - are placed at genuine physical risk by movements taught without the knowledge to modify them safely.

To protect their clients, many studios now maintain libraries of pre-approved lesson plans for their freelance instructors to follow. But building this content by hand - varied, organized by level, and modified for every pathology - consumes many expensive hours of expert time every single week.

The Solution

THE PILATES SESSION is an AI-powered application that generates safe, classical-repertoire Pilates lesson plans on demand, tailored to the specific parameters of a class - the group's level, the pathologies in the room, the prop in use, and the target muscle groups.

Delivering this required a foundation that captures the classical repertoire across all of those dimensions at once — every exercise, paired with the right modifications, cues, and prop guidance for each context. That foundation is the dataset documented in this repository. (The application itself - how it recommends and generates lesson plans - is covered in the Space README.

Business Impact

• Saves expensive expert hours - automates the weekly creation of varied, leveled lesson plans that would otherwise demand senior-instructor time.
• Reduces liability and injury risk - grounds every class in the safety-validated classical repertoire, with built-in modifications for each pathology.
• Guarantees consistency and quality - ensures a baseline standard of instruction regardless of an individual instructor's training background.
• Develops instructors - the spoken-instruction playback models professional cueing and delivery for newer teachers.
• Scales effortlessly - generates fresh, varied content on demand, so studios never repeat a class or run out of material.

---

Dataset Description

Dataset at a Glance

Property	Value
Rows	13,471 teaching protocols
Columns	11 (6 categorical/text, 5 boolean)
Classical exercises	36
Difficulty levels	4
Age groups	4
Pathology conditions	11 (10 conditions + Healthy)
Props	4 (including "No Prop")
Instruction length	~49–222 words per entry

Data Description

Each row in the dataset is a single, fully-contextualized Pilates teaching protocol. Rather than describing an exercise in the abstract, every entry places one classical mat exercise into a specific teaching context — a difficulty level, an age group, a pathology to accommodate, and a prop to use — and pairs it with a detailed, structured instruction for how to teach it safely in that exact context.

This means the dataset is fundamentally combinatorial: the 36 classical exercises are expanded across every meaningful combination of level × age group × pathology × prop, producing thousands of distinct, context-aware protocols. A single exercise like The Hundred therefore appears many times — once as a beginner protocol for a healthy teen with no prop, once as an advanced protocol for an adult with disc herniation using a Swiss Ball, and so on. This structure is what allows the downstream application to retrieve exactly the right protocol for the specific parameters of a class.

The five boolean muscle-group columns tag which regions of the body each exercise targets, enabling the application to build lessons around a chosen muscle focus.

How the Dataset Was Generated

The dataset was generated synthetically using a pre-trained large language model (Mistral-7B-Instruct) combined with a controlled, context-grounded prompting strategy. Rather than asking the model for generic content, each of the 13,471 teaching protocols was produced by injecting precise, structured parameters into a carefully engineered prompt — a technique that keeps every generation accurate, on-format, and clinically grounded.

The prompting strategy. Each generation used a strict system prompt that fixed the model's role (a certified Pilates instructor writing teaching notes) and enforced a rigid output structure — Setup, Breath, Cue, Safety, Prop — with explicit rules (no direct address, no motivational filler, base the answer only on the given parameters). A per-row user prompt then injected the exact context for that protocol: the exercise's standard execution and cues, the difficulty-specific variation, the pathology-specific safety modification, and the prop placement. This context-injection approach is what makes each of the thousands of outputs distinct, controlled, and safe.

Grounding the prompts. To supply that context, the prompts drew on a structured clinical knowledge base (a JSON file encoding all 36 classical exercises with their execution, cues, difficulty variations, prop options, and per-pathology modifications). A combination engine paired every exercise with each valid combination of level, pathology, prop, and age group, filtering out clinically invalid pairings (inapplicable props, contraindicated exercises, age-restricted movements) before prompting. The five boolean muscle-group tags were then derived programmatically from each exercise's primary muscles.

The result is a large, consistent, and clinically-grounded text dataset where the core generated content — the full teaching protocol in instruction_and_Breathing — was authored by the language model under tightly controlled prompting.

Data Dictionary

Column	Type	Description
Classical_Exercise_Name	string	The classical Pilates mat exercise (36 unique, e.g. The Hundred, Spine Twist, Teaser I).
Difficulty_Level	string	Difficulty of the exercise variation: Beginner, Intermediate, Advanced, Super Advanced.
Age_Group	string	Target age group: Teen, Young Adult, Adult, Third Age.
Pathology_Condition	string	The health condition the protocol is adapted for: Healthy, Disc Herniation, Frozen Shoulder, Hip OA, Osteoporosis, Pregnancy T1, Pregnancy T2-T3, Knee Pain, Cervical Strain, Scoliosis, wrist_sensitivity.
Prop	string	Equipment used in the protocol: No Prop, Magic Circle, Swiss Ball, Pole/Stick.
Is_Core	boolean	Whether the exercise targets the core / powerhouse.
Is_Spine	boolean	Whether the exercise targets the spine / back extensors.
Is_Lower_Body	boolean	Whether the exercise targets the lower body (hips, glutes, legs).
Is_Shoulder	boolean	Whether the exercise targets the shoulder girdle.
Is_Neck	boolean	Whether the exercise targets the cervical / neck region.
instruction_and_Breathing	string	The full structured teaching protocol, organized into Setup, Breath, Cue, Safety, and Prop sections.

---

Exploratory Data Analysis (EDA)

Step 1 - Raw Data Quality Audit

Because the dataset was authored by a language model, there is no guarantee that every entry was generated cleanly or consistently. The EDA therefore begins with a deliberate, read-only audit of the raw generated output — 13,627 rows, before any cleaning — whose sole purpose is to diagnose problems, not yet fix them. This "inspect first, clean second" separation is what lets us trust the dataset the application ultimately relies on.

Structure & missing values

• The dataset is structurally sound: 11 columns across two data types — 6 text (object) and 5 boolean.
• Only a single column contains missing values — Prop, with 5,531 nulls; every other column is fully populated.
• These nulls are not true missing data: a null Prop simply encodes "no prop used" — a semantic blank rather than a gap — and is handled in the cleaning step that follows.

Categorical distributions (clean, with no unexpected or malformed values)

• Difficulty_Level — four levels, almost perfectly balanced (~3,400 rows each).
• Age_Group — Teen, Young Adult, and Adult are evenly represented; Third Age is intentionally smaller, since only a curated subset of exercises is appropriate for that population.
• Pathology_Condition — all eleven conditions are well represented, led by Healthy and tapering gradually toward the pregnancy-specific protocols.
• Prop — four values (including the "no prop" blanks), led by no-prop, then Magic Circle, Swiss Ball, and Pole/Stick.

Text & logic integrity

• Zero unusually short instructions — every protocol clears the 25-word threshold.
• Zero rows containing broken literal-newline artifacts — a common LLM output glitch, absent here.
• Zero exercises left without a muscle-group tag — every protocol maps to at least one region of the body.
• Consistent instruction format — a diverse manual sample of generated protocols confirmed that every instruction reliably follows the exact structure defined in the prompt (Setup, Breath, Cue, Safety, Prop), demonstrating that the model faithfully adhered to the system prompt and produced clean, uniformly-formatted output.

The one genuine issue

• 156 duplicate rows — the single real problem surfaced by the audit, resolved in the cleaning step.

Together, these checks confirm that the generation pipeline produced a consistent, well-formed dataset, with only two items to address downstream: normalizing the semantic blanks in Prop, and removing the 156 duplicates.

---

Step 2 - Data Cleaning

With the audit complete, the cleaning step addressed the two issues it surfaced. First, the semantic blanks in the Prop column were normalized: the nulls representing "no prop used" were replaced with an explicit No Prop value, and a follow-up check confirmed 0 missing values across the entire dataset. Second, the 156 duplicate rows identified in the audit were removed. The result is the final, clean dataset used throughout the rest of the project:

cleaned dataset: 13,471 rows.

---

Step 3 - Instruction Length Analysis

The richest column in the dataset is the generated instruction_and_Breathing text, so the analysis begins there. Instruction length (in words) is a useful proxy for how much teaching content each protocol carries — and, because the text was model-generated, its distribution is also a window into how consistently the generation behaved across thousands of different contexts.

---

Distribution & outliers

The box plot shows a tight, well-behaved distribution. The vast majority of protocols cluster in a narrow band — roughly 135–165 words — around a median of about 150. The whiskers reach up toward ~215 words, and the only outliers sit at the short end: a small handful of protocols around 49–55 words. This tightness is itself a finding — despite being generated across thousands of different parameter combinations, the model produced instructions of remarkably consistent length, a strong signal that the generation pipeline was stable rather than erratic.

---

Edge cases

Inspecting the extremes confirms that even the outliers are legitimate, fully-formed protocols rather than generation failures. The longest instruction (222 words) belongs to Single Leg Circles, and the shortest (49 words) to Open Leg Rocker — both complete, correctly structured, and readable. No truncated, empty, or broken generations hide at the tails.

---

Length across age, difficulty, and pathology

Breaking the length distribution down by category reveals where the (small) variation actually comes from:

• Age Group — the four curves overlap heavily and peak at a similar height, meaning instruction length is largely independent of age. The one nuance: Third Age peaks slightly further right — its most common instruction length is a touch longer, consistent with the extra care and cueing these protocols carry for an older population.
• Difficulty Level — the levels also overlap closely, but Intermediate rises to a noticeably higher, narrower peak than the others, indicating its instruction lengths are more tightly concentrated (lower variance) around the typical value.
• Pathology Status — the most meaningful split. The Healthy curve is tall and narrow — lengths are concentrated and predictable — while the Pathology curve is broader and lower, spreading toward longer instructions. Protocols for clients with a pathology are therefore both longer on average and more variable in length — exactly what we would expect if the model is adding substantive, condition-specific safety content on top of the standard instruction.

---

Average length by exercise × pathology

The heatmap makes that pattern concrete, cross-tabulating average instruction length for every exercise against every pathology.Two readings stand out:

• Darker rows — a handful of exercises carry longer instructions across the board, regardless of pathology: Single Leg Circles, Single Leg Stretch, Spine Stretch Forward, and The Swan (Swan Dive). These tend to be movements with more complex setup and cueing.
• Darker columns — more tellingly, certain pathologies pull instructions longer across almost every exercise: Scoliosis, Cervical Strain, Disc Herniation, and Hip OA. These are precisely the conditions that demand the most careful modification and contraindication guidance — so the extra length is not noise; it is the model correctly investing more words in safety exactly where it matters most.

---

What this tells us

Instruction length is consistent enough to confirm a stable, reliable generation process, yet the variation that does exist is clinically meaningful rather than random. The dataset spends its words where they count — on the pathologies and exercises that genuinely require more detailed, safety-oriented instruction. This is an early but strong indication that the generated content is not merely well-formatted, but substantively responsive to the clinical context of each protocol.

---

Step 4 - Categorical & Structural Patterns

Beyond instruction length, this stage looks at the shape of the dataset itself — how its categories are distributed, how the exercises map to clinical safety and to muscle groups, and what vocabulary dominates the generated text.

---

Distribution of Difficulty, Prop, and Age Group

Three pie charts summarize the core categorical balance of the dataset:

• Difficulty Level is perfectly even — each of the four levels accounts for 25% of the data, by design.
• Prop is led by No Prop (≈40%), followed by Magic Circle (≈25%), Swiss Ball (≈20%), and Pole/Stick (≈15%) — a realistic spread that reflects how often each apparatus genuinely applies across the repertoire.
• Age Group is balanced across Teen, Young Adult, and Adult (≈28.7% each), with Third Age intentionally smaller (≈14%), since only a curated, appropriate subset of exercises is offered to that population.

This confirms a well-proportioned dataset with no accidental over-representation skewing the downstream engine.

---

Pathology Status per Exercise

Each protocol's instruction text was classified by safety status. In practice the data resolves into two dominant categories — SAFE (the large majority) and CONTRAINDICATED — revealing, for every exercise, across how many pathologies it remains safe versus how many it is unsafe for.

• Most of every bar is SAFE: for the majority of pathologies, most exercises can be taught safely.
• The CONTRAINDICATED share varies sharply by exercise. The most contraindicated — unsafe across the most pathologies — are the dynamic, high-load movements: Double Leg Stretch, Jack Knife, Roll Over, and Single Straight Leg Stretch (Scissors).
• At the opposite end, several exercises are safe for virtually every pathology: Shoulder Bridge, Spine Twist, and the entire Side Kick series (Front/Back, Inner Thigh, Small Circles, Up/Down) — the gentler, more controlled movements.

This is a clinically coherent result: the exercises the model flags as unsafe most often are exactly those that demand the greatest spinal load and control, while the low-risk, stabilizing movements remain broadly accessible. The dataset's safety logic behaves the way an expert instructor would expect.

---

Muscle Group Coverage per Exercise

This heatmap maps which of the five muscle regions each exercise engages:

• Lower Body is engaged by almost every exercise (≈90%), and Core by the large majority (≈72%) — unsurprising, since Pilates is built around the powerhouse and full-body integration.
• Spine is mixed (≈50%), and Shoulder is engaged less often (≈38%).
• Neck is the rare tag (≈1%): the row is almost entirely empty, lit up by only a couple of exercises such as Neck Roll. Crucially, no exercise engages all five regions, and Neck almost never co-occurs — a realistic reflection of how the classical repertoire actually distributes its emphasis.

---

Vocabulary of the Instructions

A word cloud of the full instruction corpus provides a qualitative check on what the model actually wrote. The dominant terms are purely anatomical and directional — spine, mat, arms, legs, hip — surrounded by movement and control language: powerhouse, control, throughout, rotation, extended, supine, pelvis, breath, safety. What is striking is what is absent: there is no motivational or conversational filler. The vocabulary is exactly the clean, technical, instructor-facing register the system prompt demanded — a direct confirmation that the generation adhered to its intended style.

---

What this tells us

Taken together, these views show a dataset that is not only well-balanced categorically, but internally coherent in a domain-meaningful way. Its safety classifications track real biomechanical risk, its muscle-group coverage mirrors the true emphasis of the classical method, and its language stays disciplined and clinical throughout. The structure the model produced is consistent with genuine Pilates expertise — not just superficially well-formatted output.

---

Step 5 - Statistical Summary

The final stage steps back from individual plots to summarize the dataset as a whole. Rather than examining one feature at a time, it asks two broader questions: what does each column look like statistically — its range, its most common values, its diversity — and how do the columns relate to one another? Together, a descriptive summary of the categorical and text fields, a coverage breakdown of the muscle-group flags, and a correlation analysis across all variables give a final, high-level verdict on the quality and balance of the data before it is put to use.

Descriptive statistics - categorical & text columns

Statistic	Classical_Exercise_Name	Difficulty_Level	Age_Group	Pathology_Condition	Prop	instruction_and_Breathing
count	13,471	13,471	13,471	13,471	13,471	13,471
unique	36	4	4	11	4	3,825
top	Spine Twist	Super Advanced	Teen	Healthy	No Prop	• Setup: Side-lying position with the top leg ...
freq	704	3,373	3,867	1,373	5,379	12

The summary reinforces what the earlier plots showed, and adds a few useful details. Every column is fully populated (count = 13,471 throughout). The categorical columns confirm the expected cardinality — 36 exercises, 4 difficulty levels, 4 age groups, 11 pathology conditions, and 4 prop values — with the most frequent value in each being a sensible, well-represented default (Spine Twist, Super Advanced, Teen, Healthy, No Prop). Most revealing is the instruction_and_Breathing column: it holds 3,825 unique instruction texts across the 13,471 rows, and the single most common instruction appears only 12 times. This is a healthy signal — no boilerplate response dominates the dataset, and the repetition that does exist is expected, since the same exercise at the same level can legitimately share instructions across combinations that don't alter the teaching content.

Muscle group coverage - share of protocols engaging each region

Muscle Group	% True
Is_Core	71.8%
Is_Spine	50.0%
Is_Lower_Body	90.1%
Is_Shoulder	38.2%
Is_Neck	1.0%

The muscle-group coverage quantifies the emphasis of the repertoire: Lower Body (90.1%) and Core (71.8%) dominate, consistent with Pilates' foundational focus on the powerhouse and lower-body integration, while Spine sits at exactly half and Shoulder at 38.2%. Neck is engaged by just 1.0% of protocols — a deliberate rarity, reflecting how few classical exercises isolate the cervical region.

Categorical correlation - Cramér's V

Because every field in this dataset is categorical or boolean — none are numeric — a standard (Pearson) correlation does not apply. Instead, association is measured with Cramér's V, a statistic built on the chi-squared test that quantifies the strength of association between two categorical variables on a convenient 0-to-1 scale (0 = independent, 1 = perfectly associated). It is the appropriate tool for exactly this kind of fully-categorical data, and the resulting heatmap tells a clear and reassuring story about how the dataset is structured:

• The parameters are largely independent of one another. Difficulty_Level is completely decoupled from everything else (0.00 across the board), and Pathology_Condition and Age_Group show only negligible associations (about 0.00–0.09) with the other fields. This is exactly what a well-designed combinatorial dataset should look like: level, pathology, and age are distributed evenly across the exercises, with no hidden bias linking them.
• The only strong correlations are deterministic and expected. Classical_Exercise_Name correlates perfectly (1.00) with all five muscle-group flags — unsurprising, since the muscle tags are derived directly from the exercise. It also shows a moderate association with Prop (0.33) and a mild one with Age_Group (0.20), reflecting that some props only apply to certain exercises and that some exercises are age-restricted.
• Muscle groups relate to each other in anatomically sensible ways — mild associations such as Spine–Shoulder (0.26) and Lower Body–Neck (0.30) capture which regions tend to be engaged together across the repertoire.

In short, there are no surprising or spurious dependencies: the strong signals are the ones that should be strong (an exercise determines its muscles), and the design parameters remain cleanly orthogonal.

---

EDA - Conclusion

The exploratory analysis confirms that the synthetic generation pipeline produced a dataset that is clean, well-balanced, and clinically coherent. Structurally, it is sound — no genuine missing data, no malformed values, and only minor cleaning required. Statistically, its parameters are evenly distributed and independent, so no dimension dominates or biases another. And most importantly, its content behaves the way real Pilates expertise would predict: instruction length grows where safety demands it, contraindications concentrate in the highest-load movements, muscle coverage mirrors the method's true emphasis, and the language stays disciplined and technical throughout. The dataset is a trustworthy foundation for the recommendation and generation engines built on top of it.

---

Data Quality & Limitations

While the exploratory analysis confirmed that the dataset is clean, well-structured, and internally coherent, it is important to be transparent about its limitations — especially given its safety-sensitive domain.

Machine-generated phrasing. The wording of all 13,471 protocols was generated by a language model. Not every individual row has been read and verified one by one, so isolated phrasing deviations may exist in specific combinations that were not manually reviewed.

Uneven coverage. Some categories are intentionally rare — Is_Neck appears in only about 1% of protocols, and the Third Age group is smaller than the others. This faithfully mirrors the real classical repertoire (few exercises isolate the neck; fewer are appropriate for older populations), but it also means the data offers fewer examples in these areas.

Limited scope. The dataset covers 36 classical mat exercises, 11 pathology conditions, and 4 props. This captures the core of the classical mat repertoire, but not the entirety of the Pilates world — it does not include apparatus work (such as the Reformer or Cadillac), nor every possible pathology or population.

Not a substitute for professional judgment. Although the dataset reaches a high clinical quality, it is intended as a planning aid, not a guarantee of physical safety. Responsibility for the physical safety of any class remains with the supervising instructor, and exercises and combinations should be independently reviewed before real-world use. The author does not assume liability for the physical outcomes of applying this content.

---

Tech Stack & Environment

To view the complete data generation and analysis pipeline, open the project notebook (.ipynb) below:

The dataset generation and exploratory analysis were performed using the following Python environment:

# System and utility
import os, re, random, warnings, time, json, itertools
from pathlib import Path
from collections import Counter

# Data manipulation
import numpy as np
import pandas as pd
import scipy.stats as ss

# Synthetic data generation (LLM)
import torch
from transformers import pipeline
from tqdm import tqdm

# Visualization
import matplotlib.pyplot as plt
import plotly.express as px
import plotly.graph_objects as go
from plotly.subplots import make_subplots
from wordcloud import WordCloud, STOPWORDS
from PIL import Image, ImageDraw

---

Lia Prop | July 2026