prismer-ai/ml-experiment

ML Experiment Skill

Description

Design, implement, and evaluate machine learning experiments with reproducible workflows, proper baselines, and statistical analysis.

Tools Used

• jupyter_execute - Execute ML code in Python (auto-switches to Jupyter)
• jupyter_notebook - Manage experiment notebooks
• update_notebook - Set up experiment cells
• update_latex - Write experiment results to papers
• latex_compile - Compile CS conference papers (auto-switches to LaTeX)
• arxiv_to_prompt - Read related work from arXiv papers
• update_notes - Write experiment logs and analysis summaries

Capabilities

Experiment Design

• Proper train/validation/test splits
• Cross-validation and bootstrap confidence intervals
• Ablation study design
• Hyperparameter search (grid, random, Bayesian)

Implementation

• PyTorch and TensorFlow model building
• Data loading and augmentation pipelines
• Training loops with logging and checkpointing
• Distributed training setup

Evaluation

• Standard metrics per task (accuracy, F1, BLEU, mAP, etc.)
• Statistical significance testing (paired t-test, bootstrap)
• Comparison with baselines
• Error analysis and visualization

Usage Patterns

Run an Experiment

When user says: "Train a model for [task]"

1. Clarify dataset, metrics, and baselines
2. Implement data loading and preprocessing
3. Build model architecture
4. Train with proper logging
5. Evaluate and compare to baselines
6. Report results with confidence intervals

Reproduce a Paper

When user says: "Reproduce [paper title/arXiv ID]"

1. Fetch paper using arxiv_to_prompt
2. Extract key method details
3. Implement core algorithm
4. Run experiments matching paper setup
5. Compare results to reported numbers

Tool Examples

Train and evaluate a classifier

# via jupyter_execute
import torch
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
# ... train model ...
print(classification_report(y_test, predictions))

Run ablation study

# via jupyter_execute
configs = [
    {"name": "full", "use_augmentation": True, "use_dropout": True},
    {"name": "no_aug", "use_augmentation": False, "use_dropout": True},
    {"name": "no_dropout", "use_augmentation": True, "use_dropout": False},
]
results = {c["name"]: train_and_eval(**c) for c in configs}

Validation checkpoints

• Verify data shapes match expected dimensions before training
• Check that loss is decreasing after the first few epochs
• Confirm test set has no overlap with training data