Versioning Models: Why Filenames and Metadata Matter
Learners will design a naming convention that embeds training date, dataset version, and metric score, and write a JSON metadata sidecar for governance.
Versioning Models: Why Filenames and Metadata Matter is a free Machine Learning Academy lesson on CoddyKit — lesson 2 of 4. You can read the complete lesson below for free — then practise it hands-on in the browser with a built-in code editor and a 24/7 AI tutor. It is part of the Machine Learning Academy learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.
The Problem Without Versioning
Without a disciplined versioning strategy, teams quickly accumulate model.pkl, model_v2.pkl, model_final.pkl, model_FINAL_v2.pkl files with no record of which was trained on what data, which metric it achieved, or which is actually running in production. This chaos leads to deploying stale models, losing the best checkpoint, or being unable to reproduce a past result for debugging.
What to Encode in the Filename
A good model filename should encode enough context to be self-describing: dataset, algorithm, date, and optionally the primary metric and the dataset version or git SHA. This makes the model registry folder a readable audit trail at a glance, without needing to open each file.
from datetime import date
def model_filename(dataset, algorithm, metric_name, metric_value,
data_version='v1', ext='joblib'):
today = date.today().strftime('%Y%m%d')
metric_str = f'{metric_name}{int(metric_value * 100)}'
return f'{dataset}__{algorithm}__{today}__dv{data_version}__{metric_str}.{ext}'
# Examples
print(model_filename('titanic', 'rf', 'acc', 0.834, data_version='2'))
print(model_filename('fraud', 'xgb', 'auc', 0.971))
print(model_filename('cancer', 'logreg', 'f1', 0.955))The JSON Metadata Sidecar
Every model file should have a companion JSON sidecar file with the same stem name. The sidecar documents everything the filename cannot: library versions, hyperparameters, training set size, test set performance across all metrics, feature list, and any notes about the training run. This is the minimum viable model card.
import json
import sklearn, sys
from datetime import datetime
def save_metadata(path, dataset, algorithm, params, metrics, features, notes=''):
meta = {
'model_path': path,
'dataset': dataset,
'algorithm': algorithm,
'hyperparameters': params,
'metrics': metrics,
'features': features,
'sklearn_version': sklearn.__version__,
'python_version': sys.version.split()[0],
'trained_at': datetime.utcnow().isoformat(),
'notes': notes
}
meta_path = path.replace('.joblib', '_metadata.json')
with open(meta_path, 'w') as f:
json.dump(meta, f, indent=2)
print('Metadata saved to:', meta_path)
return meta
# Usage example
save_metadata(
'/tmp/cancer__logreg__20260620__dv1__acc97.joblib',
dataset='breast_cancer', algorithm='LogisticRegression',
params={'C': 1.0, 'max_iter': 300},
metrics={'accuracy': 0.9789, 'roc_auc': 0.9941, 'f1': 0.9831},
features=['mean radius', 'mean texture', '... 30 total'],
notes='Trained on full UCI breast cancer dataset'
)Dataset Version Tracking
The training dataset itself must be versioned. A model trained on data_v1.csv and another on data_v2.csv should never have the same model identifier. Options for dataset versioning: store a Git SHA of the data file, record an MD5/SHA256 hash of the CSV, or use a data versioning tool like DVC (Data Version Control) which manages dataset lineage the same way Git manages code.
import hashlib
def file_hash(path, algo='sha256'):
h = hashlib.new(algo)
with open(path, 'rb') as f:
for chunk in iter(lambda: f.read(65536), b''):
h.update(chunk)
return h.hexdigest()[:12] # first 12 hex chars as short ID
# Example: hash the model file itself as a unique ID
model_path = '/tmp/cancer_model.joblib'
model_hash = file_hash(model_path)
print('Model hash (short):', model_hash)Semantic Versioning for Models
Borrow from software engineering: use semantic versioning (MAJOR.MINOR.PATCH) for models. MAJOR: breaking change (different feature set or incompatible schema). MINOR: performance improvement with same API. PATCH: bug fix or minor recalibration. This convention helps downstream consumers understand the impact of updating their dependency on the model.
model_registry = [
{'version': '1.0.0', 'algorithm': 'LogisticRegression', 'auc': 0.921,
'note': 'Initial production model'},
{'version': '1.1.0', 'algorithm': 'LogisticRegression', 'auc': 0.935,
'note': 'Retrained on 3 months more data'},
{'version': '2.0.0', 'algorithm': 'XGBoost', 'auc': 0.971,
'note': 'New algorithm; feature set changed — incompatible schema'}
]
print('Model Registry:')
for entry in model_registry:
print(f" v{entry['version']} AUC={entry['auc']} {entry['note']}")A Simple Local Model Registry
A lightweight model registry can be a directory with a registry.json file that indexes all saved models. Each entry records the filename, version, key metrics, and the production flag. The deployment script reads this file to determine which model to load.
import json
import os
REGISTRY_PATH = '/tmp/model_registry.json'
def register_model(filename, version, metrics, is_production=False):
try:
with open(REGISTRY_PATH) as f:
registry = json.load(f)
except FileNotFoundError:
registry = []
# Mark all as not-production if this one is production
if is_production:
for entry in registry:
entry['is_production'] = False
registry.append({
'filename': filename,
'version': version,
'metrics': metrics,
'is_production': is_production
})
with open(REGISTRY_PATH, 'w') as f:
json.dump(registry, f, indent=2)
print(f'Registered v{version} (production={is_production})')
register_model('cancer__logreg__v1.0.0.joblib', '1.0.0',
{'accuracy': 0.979, 'auc': 0.994}, is_production=True)Reading the Production Model from Registry
At inference time, the service loads the registry and finds the model flagged as production, then loads that specific file. This decouples the deployment script from hardcoded filenames — updating the production model only requires updating the registry flag, not modifying serving code.
import json
import joblib
def load_production_model(registry_path, model_dir='/tmp'):
with open(registry_path) as f:
registry = json.load(f)
prod = next((r for r in registry if r['is_production']), None)
if prod is None:
raise RuntimeError('No production model registered!')
path = f'{model_dir}/{prod["filename"]}'
print(f'Loading production model: {prod["filename"]} (v{prod["version"]})')
print(f'Metrics: {prod["metrics"]}')
# return joblib.load(path) # would load for real
return None # demo
load_production_model('/tmp/model_registry.json')MLflow: Professional Model Registry
MLflow is the industry-standard tool for experiment tracking and model registry. It logs parameters, metrics, and artefacts for each training run; lets you compare runs via a UI; and provides a Model Registry with staging/production/archived states. For teams, MLflow replaces manual JSON registries with a robust, queryable backend.
import mlflow
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris
from sklearn.model_selection import cross_val_score
import numpy as np
X, y = load_iris(return_X_y=True)
with mlflow.start_run(run_name='logreg_iris_v1'):
model = LogisticRegression(C=1.0, max_iter=200)
cv_score = cross_val_score(model, X, y, cv=5).mean()
mlflow.log_param('C', 1.0)
mlflow.log_param('max_iter', 200)
mlflow.log_metric('cv_accuracy', cv_score)
model.fit(X, y)
mlflow.sklearn.log_model(model, 'model')
print(f'CV Accuracy: {cv_score:.4f}')
print('Run logged to MLflow')Tagging Model Versions
Tags add freeform key-value annotations to a model version — useful for recording experiment context, the analyst's name, the task type, or whether the model passed a fairness audit. Tags are queryable, making it easy to filter the registry by any attribute.
# Simulated registry entry with tags (no MLflow required)
model_entry = {
'version': '2.1.0',
'filename': 'fraud__xgb__2.1.0.joblib',
'tags': {
'analyst': 'data-science-team',
'task': 'binary-classification',
'fairness_audit': 'passed',
'retrain_trigger': 'monthly-schedule',
'deployment_region': 'eu-west-1'
},
'metrics': {'roc_auc': 0.971, 'precision': 0.83, 'recall': 0.79}
}
print('Model entry:')
print(json.dumps(model_entry, indent=2))Automated Promotion Criteria
Define clear promotion criteria before any model goes to production: the new model must achieve at least X% AUC improvement, must pass a fairness check, must not degrade on any monitored demographic slice, and must complete inference within Y milliseconds. Encoding these criteria in code (not documentation) lets a CI/CD pipeline automate promotion decisions objectively.
def should_promote(new_metrics, baseline_metrics, min_auc_improvement=0.005):
if new_metrics['auc'] < baseline_metrics['auc'] + min_auc_improvement:
return False, 'AUC improvement too small'
if new_metrics.get('fairness_delta', 0) > 0.05:
return False, 'Fairness constraint violated'
if new_metrics.get('latency_ms', 0) > 100:
return False, 'Latency too high'
return True, 'All criteria met'
baseline = {'auc': 0.921}
candidate = {'auc': 0.937, 'fairness_delta': 0.02, 'latency_ms': 45}
promote, reason = should_promote(candidate, baseline)
print(f'Promote: {promote} — {reason}')Model Cards and Documentation
Beyond technical metadata, a model card (coined by Google) documents the model's intended use, limitations, training data characteristics, performance across demographic subgroups, and ethical considerations. For models making consequential decisions (loan approvals, medical triage), model cards are becoming a regulatory requirement. Include at minimum: intended use, out-of-scope uses, performance metrics, and known failure modes.
Quick Check
Test your understanding of model versioning and metadata from this lesson.
Lesson Recap
In this lesson you learned: descriptive filenames embedding dataset, algorithm, date, and metric make your model directory self-documenting, JSON metadata sidecars record library versions, hyperparameters, and metrics needed for governance and reproducibility, and a model registry (local JSON or MLflow) decouples serving code from hardcoded filenames. Next up we wrap a saved model in a FastAPI endpoint to serve predictions over HTTP.
Frequently asked questions
Is the “Versioning Models: Why Filenames and Metadata Matter” lesson free?
Yes — the full text of “Versioning Models: Why Filenames and Metadata Matter” is free to read here on the web, and the Machine Learning Academy course includes 4 lessons in total. To practise it interactively (a built-in code editor and a 24/7 AI tutor) and unlock the rest of the Machine Learning Academy course, upgrade to CoddyKit PRO.
What will I learn in “Versioning Models: Why Filenames and Metadata Matter”?
Learners will design a naming convention that embeds training date, dataset version, and metric score, and write a JSON metadata sidecar for governance. You practise Machine Learning Academy with hands-on code you run directly in the browser, and a 24/7 AI tutor answers your questions as you work through the lesson.
Do I need any experience to start Machine Learning Academy?
No prior experience is required. Machine Learning Academy on CoddyKit is structured for beginners through advanced learners; this is — lesson 2 of 4, so you can start here or from the beginning and move at your own pace.
How long does the “Versioning Models: Why Filenames and Metadata Matter” lesson take?
Most CoddyKit lessons take about 5–10 minutes. Each one is bite-sized and interactive, so you make steady progress and pick up exactly where you left off across the web and the app.
Can I write and run code in this Machine Learning Academy lesson?
Yes. Every Machine Learning Academy lesson includes a built-in code editor, so you write and run real code right in your browser and get instant AI feedback — no local setup required.
All lessons in this course
- Saving Models with joblib and pickle
- Versioning Models: Why Filenames and Metadata Matter
- Serving Predictions with a FastAPI Endpoint
- Monitoring Predictions: Logging Inputs and Outputs