File size: 5,787 Bytes
0dc4d1a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 | ---
pretty_name: Learning Paths in Intro Mechanics (sample)
license: other
license_name: creative-data-lab-dataset-license
license_link: https://creative-data-lab-production.up.railway.app/terms.html
language:
- en
task_categories:
- question-answering
- text-generation
tags:
- education
- tutoring
- synthetic
- misconceptions
- creative-data-lab
- physics
size_categories:
- n<1K
configs:
- config_name: master
data_files:
- split: train
path: master/records.jsonl
default: true
- config_name: sft
data_files:
- split: train
path: exports/sft.jsonl
- config_name: tutor
data_files:
- split: train
path: exports/tutor.jsonl
- config_name: preference
data_files:
- split: train
path: exports/preference.jsonl
- config_name: step_labels
data_files:
- split: train
path: exports/step_labels.jsonl
---
# Learning Paths in Intro Mechanics: free sample
*Creative Data Lab · 8 of 45 records from batch-001*
Learning paths through introductory mechanics. Each record is a problem, one or two realistic wrong attempts
tagged with a documented misconception, Socratic tutor feedback, and a correct worked solution.
This is a **free sample** so you can check the schema and quality. The full dataset is
[creativedatalab/physics-v1](https://huggingface.co/datasets/creativedatalab/physics-v1) (US$149 per organization, instant access after purchase at
[creative-data-lab-production.up.railway.app](https://creative-data-lab-production.up.railway.app/#datasets)).
## Load it
```python
from datasets import load_dataset
records = load_dataset("creativedatalab/physics-v1-sample", "master") # full records (default)
sft = load_dataset("creativedatalab/physics-v1-sample", "sft") # problem -> worked solution
tutor = load_dataset("creativedatalab/physics-v1-sample", "tutor") # problem + wrong attempt -> guiding question
preference = load_dataset("creativedatalab/physics-v1-sample", "preference") # chosen / rejected pairs
steps = load_dataset("creativedatalab/physics-v1-sample", "step_labels") # per-step labels, first error marked
```
## Files
| File | Sample rows | Full dataset |
|---|---:|---:|
| `master/records.jsonl` | 8 | 45 |
| `exports/sft.jsonl` | 8 | 45 |
| `exports/tutor.jsonl` | 8 | 49 |
| `exports/preference.jsonl` | 8 | 49 |
| `exports/step_labels.jsonl` | 16 | 94 |
## How it was made
- **Ground truth is computed by code.** Every problem comes from a parameterised template; the correct answer
and each misconception's wrong answer are computed with sympy, never by a language model.
- **Text is written by an open-weight model**: `mlx-community/Qwen2.5-32B-Instruct-4bit` (Apache 2.0), which writes problem wording,
student attempts, feedback and solution steps around the computed values.
- **Every record passed automated checks**: answer matches truth within 1%; each wrong attempt lands exactly on its
misconception's predicted answer; feedback does not leak the answer; all arithmetic in the text recomputes;
no LaTeX or formatting debris; no duplicate problems.
- **Human review**: 0 of 53 records reviewed so far.
Batch yield: 53 of 56 problems kept (45 on the first try).
## Coverage
| Topic | Records |
|---|---:|
| E Work and energy | 10 |
| K Kinematics | 15 |
| N Newton's laws | 20 |
| Rung | Records |
|---|---:|
| R1 Intuition | 11 |
| R2 Definition | 4 |
| R3 One step | 24 |
| R4 Multi step | 6 |
| Concept | Records |
|---|---:|
| E1 Work by a force | 4 |
| E2 Kinetic energy / work–energy theorem | 4 |
| E5 Energy with friction | 2 |
| K1 Displacement vs. distance | 4 |
| K6 Free fall | 7 |
| K7 Projectile motion | 4 |
| N4 Third law (interaction pairs) | 4 |
| N5 Weight | 4 |
| N6 Friction (static and kinetic) | 4 |
| N7 Inclined planes | 4 |
| N8 Tension and pulleys (Atwood) | 4 |
| Misconception | Records |
|---|---:|
| MC01 Heavier objects fall faster | 3 |
| MC02 Motion requires a force in its direction | 4 |
| MC03 Zero velocity means zero acceleration | 4 |
| MC04 Distance equals displacement | 4 |
| MC09 Bigger object exerts bigger force | 4 |
| MC10 Normal force always equals mg | 4 |
| MC11 Static friction always equals μs·N | 4 |
| MC13 Horizontal speed changes fall time | 4 |
| MC14 Kinetic energy is linear in speed | 4 |
| MC16 Mechanical energy conserved despite friction | 2 |
| MC17 Perpendicular forces do work | 4 |
| MC18 Mass and weight are the same | 4 |
| MC19 Rope tension equals hanging weight | 4 |
## Record schema (master)
`id`, `split`, `topic`, `concept_id`, `prereqs`, `rung` (R1 intuition → R5 transfer), `difficulty`,
`problem` (template, params, text, options), `ground_truth` (answer, unit, tolerance), `attempts`
(misconception_id, text, final_answer), `feedback`, `solution` (steps, final_answer), `provenance`
(generator, template version, seed, check results, human review).
## Step labels: how to read them
Correct solutions are labelled +1 at every step. For misconception attempts, the first step that reaches the
misconception's wrong answer, and every step after it, are labelled -1. Earlier steps are `null`: they are not
verified either way.
## Limitations
- Sample batch: 11 of 22 planned concepts, 45 training records.
- Multiple-choice stems are template-written, so wording repeats across records of the same template.
- Physics is checked by code; pedagogy (realism of attempts, quality of feedback) relies on human review.
## License
Problems are original, generated from in-house templates. Dataset license: per client agreement.
## License
Sample: [Creative Data Lab Dataset License](https://creative-data-lab-production.up.railway.app/terms.html), evaluation use. The full dataset
carries the same license for training and evaluation inside one organization.
|