smol-training-playbook

Provides a web-based UI for constructing and visualizing model training configurations without writing code. Users select hyperparameters, dataset sizes, compute resources, and training objectives through form controls that generate reproducible training scripts. The interface validates parameter combinations against known constraints and displays estimated training time and resource requirements based on model size and dataset scale.

Converts user-selected training parameters into executable Python scripts by applying parameter values to pre-built training templates. The system maintains a library of template scripts for different training paradigms (supervised fine-tuning, instruction tuning, reinforcement learning from human feedback) and injects selected hyperparameters, model identifiers, and dataset paths into template placeholders. Generated scripts are syntactically valid and immediately executable with minimal modification.

Analyzes selected model size, dataset dimensions, and hyperparameters to estimate GPU memory requirements, training duration, and computational cost. The calculator uses empirical scaling laws and hardware specifications to project resource consumption before training begins. Estimates account for batch size, sequence length, gradient accumulation, and mixed-precision training settings, displaying results in human-readable formats (GB, hours, USD).

Validates user-selected hyperparameter combinations against known constraints and best practices before script generation. The validator checks for incompatible settings (e.g., learning rate too high for model size), warns about suboptimal configurations, and suggests corrections based on training literature and empirical results. Validation rules are encoded as constraint definitions that compare parameter values against thresholds and interdependencies.

Generates comprehensive training documentation and playbooks based on selected configurations, including setup instructions, execution steps, troubleshooting guides, and expected outcomes. The documentation system creates markdown or HTML output that explains the training approach, hyperparameter rationale, and how to interpret results. Documentation is templated and customized with user selections, providing context-specific guidance rather than generic instructions.

Provides browsable catalogs of pre-trained models and datasets integrated with HuggingFace Hub, allowing users to search, filter, and preview options before selecting them for training. The interface displays model metadata (parameter count, training data, performance benchmarks), dataset statistics (size, languages, domains), and compatibility information. Selection is context-aware, suggesting compatible models and datasets based on training objective and available resources.

Orchestrates the complete training workflow from configuration through script generation and execution guidance, managing state and dependencies across steps. The system tracks configuration selections, validates constraints, generates scripts, estimates resources, and produces documentation in a coordinated pipeline. Workflow state is maintained across user sessions, allowing users to save, modify, and reuse configurations. Integration points include HuggingFace Hub APIs for model/dataset discovery and external execution environments for script running.