RPMeta — RPM Build Duration Prediction Models

Pre-trained models for RPMeta, a service that predicts RPM package build durations based on package metadata and hardware resources.

Repository Structure

lightgbm/
  native_model.txt      # LightGBM native model
xgboost/
  native_model.ubj      # XGBoost native model (universal binary JSON)
categories.json         # Categorical feature encoding map (shared by both models)

Usage

With RPMeta CLI

dnf copr enable @copr/rpmeta
dnf install rpmeta rpmeta+server

# Download the model
huggingface-cli download fedora-copr/rpmeta --local-dir /var/lib/rpmeta/models

And run the rpmeta service.

Direct download (no authentication needed)

# LightGBM model
curl -L https://huggingface.co/fedora-copr/rpmeta/resolve/main/lightgbm/native_model.txt -o native_model.txt

# XGBoost model
curl -L https://huggingface.co/fedora-copr/rpmeta/resolve/main/xgboost/native_model.ubj -o native_model.ubj

# Categories
curl -L https://huggingface.co/fedora-copr/rpmeta/resolve/main/categories.json -o categories.json

Model Details

• Training data: Copr build records from Fedora infrastructure
• Task: Regression — predict build duration in seconds from package metadata and hardware info

Versioning

Model versions are tracked via git tags (e.g. v2025.12.24). To pin a specific version:

curl -L https://huggingface.co/fedora-copr/rpmeta/resolve/v2025.12.24/lightgbm/native_model.txt -o native_model.txt

Training history and hyperparameter tuning results are preserved on the training-history branch.