juancavallotti/climbing-goal-classifier

Climbing Goal Classification Dataset

A text classification dataset for interpreting climbing training goals, designed to handle real-world ambiguity rather than idealized inputs.

The dataset supports identifying whether a goal is:

• actually about climbing
• missing key information (timeline and/or metric)
• or fully specified and actionable

This dataset was built using a mix of synthetic generation and human-in-the-loop editing to avoid template memorization and encourage semantic generalization.

---

Task

Multi-class text classification

Given a short free-form text input (e.g. a user-entered goal), predict its intent class.

Example inputs:

• reach v7 by mid year
• climb more slabs
• send V6
• by June improve technique
• peel potatoes

---

Labels

The dataset uses the following label taxonomy:

Label	Description
not_climbing	Not related to climbing at all
climbing_vague	Climbing-related intent, but no timeline or measurable metric
missing_timeline	Measurable climbing goal, but no timeline
missing_metric	Timeline present, but metric is vague or non-measurable
missing_both	Explicit goal phrasing, but missing both timeline and metric
correct	Has both a timeline and a measurable climbing metric

Notes on ambiguity

Some examples intentionally sit near label boundaries. This is by design.
Consistency and robustness were prioritized over “perfect” separability.

---

Dataset Structure

Each row contains:

• text (string): raw user input
• label (string): one of the labels above

No preprocessing (lowercasing, normalization) is applied; the text reflects realistic user input, including fragments, shorthand, casing variation, and minor noise.

---

Data Generation Process

1. Synthetic bootstrap

   • Programmatic generation of diverse climbing and non-climbing goal phrases
   • Coverage of grades, timelines, acronyms, shorthand, and variants

2. Human-in-the-loop editing

   • Manual mutation of samples to:
     • remove templates
     • introduce ambiguity
     • add fragments and edge cases
   • Targeted augmentation of weak classes (especially not_climbing)

3. Validation by failure

   • Dataset iterations were driven by model failure cases, not metrics alone
   • Perfect confusion matrices were treated as a red flag and actively broken

---

Intended Use

This dataset is well suited for:

• Goal validation and feedback systems
• On-device intent classification (e.g. CoreML)
• Assistive UX flows (“can you clarify your goal?”)
• Research on ambiguity-aware classification

It is not intended for:

• grading climbing performance
• predicting user skill
• safety-critical decision making

---

Evaluation Guidance

⚠️ Do not rely on synthetic train/test splits alone.

Recommended evaluation setup:

• small human-written holdout set (never trained on)
• macro-F1 instead of raw accuracy
• inspection of confusion between adjacent classes

Perfect accuracy usually indicates shortcut learning, not true generalization.

---

License

This dataset is released under the Apache License, Version 2.0.

You are free to:

• use the dataset for research or commercial purposes
• modify and redistribute it
• train models on it and use those models commercially

Under the conditions that you:

• include a copy of the Apache 2.0 license
• provide appropriate attribution
• clearly indicate if changes were made

The dataset is provided “as is”, without warranty of any kind.

See the LICENSE file for full terms.

---

Citation

If you use this dataset, please cite or link to the Hugging Face repository.

---

Acknowledgements

This dataset was shaped by iterative failure analysis and real-world UX constraints rather than benchmark optimization.

---

dataset_info: features: - name: text dtype: string - name: label dtype: string - name: labels dtype: class_label: names: '0': not_climbing '1': climbing_vague '2': missing_timeline '3': missing_metric '4': missing_both '5': correct splits: - name: train num_bytes: 198461 num_examples: 3032 download_size: 46297 dataset_size: 198461 configs: - config_name: default data_files: - split: train path: data/train-* license: apache-2.0 task_categories: - text-classification language: - en tags: - fitness - climbing pretty_name: Climbing Goals Detection Dataset size_categories: - 1K<n<10K