CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

Core Framework

• Artifact system with Pydantic validation for reproducible pipelines
• Support for artifact lineage tracking and metadata
• Scale factors (tiny/small/medium/full) for fast iteration
• Unix piping support between pipeline steps
• Atomic file writes for artifact metadata
• WandbTracker for optional W&B integration
• CLI generation with typer and OmegaConf

Training Recipe Structure

• Recipe package structure at src/nemotron/recipes/
• Placeholder structure for Nemotron Nano 2 recipe:
  • Stage 0: Pretraining
  • Stage 1: Instruction Tuning
  • Stage 2: Alignment
• Placeholder structure for ChipNeMo/ScaleRTL recipe:
  • Stage 0: Domain Pretraining
  • Stage 1: Supervised Fine-tuning
  • Stage 2: Reasoning Enhancement

Examples

• Training pipeline pattern examples in examples/training/:
  • hello.py - Minimal pipeline step example
  • tutorial_data_prep.py - Data preparation patterns
  • tutorial_training.py - Training loop patterns
  • tutorial_evaluation.py - Evaluation patterns

Documentation

• Comprehensive README with:
  • NVIDIA AI stack integration details (Curator, Megatron-Bridge, Automodel, NeMo-RL, Evaluator)
  • Recipe overview with paper links
  • Usage examples and quick start
• Recipe-specific READMEs with detailed TODO lists
• Contribution guidelines inspired by NVIDIA-NeMo/Megatron-Bridge

Testing

• Unit tests for Artifact system
• Test suite with pytest
• 100% test coverage for core framework

Technical Details

Dependencies:

• pydantic >= 2.0.0
• typer >= 0.12.0
• omegaconf >= 2.3.0
• rich >= 13.0.0
• wandb (optional)

Python Support: 3.10+

Package Management: uv

---

Future Releases

Planned for 0.1.0 (First Release)

Nemotron Nano 2 Recipe

• Stage 0 implementation (pretraining)
• Stage 1 implementation (instruction tuning)
• Stage 2 implementation (alignment)
• Benchmark results and validation
• Hardware requirements documentation

ChipNeMo/ScaleRTL Recipe

• Stage 0 implementation (domain pretraining)
• Stage 1 implementation (SFT)
• Stage 2 implementation (reasoning enhancement)
• RTL generation benchmarks
• Integration with test-time compute

Infrastructure

• CI/CD pipeline
• Automated testing
• Docker containers for reproducibility
• Pre-commit hooks

Documentation

• Usage guides in docs/usage/
• Deployment guides in docs/deployment/
• Video tutorials
• API documentation

---

Version History

This is the initial version of the project. Version history will be updated as releases are published.

---

Note: This project is in active development. Recipe implementations are coming soon. Contributions are welcome!