April 2026 monthly summary for metatensor/metatrain focused on reliability, training efficiency, and performance gains. Delivered targeted fixes and architectural improvements that reduce run-to-run variability, shorten training time, and improve model accuracy through better data handling and plotting of results for decision making.

February 2026 update focused on reliability improvements in the training pipeline of metatensor/metatrain. Implemented a targeted bug fix to correct num_epochs handling for restart and finetune contexts, ensuring the correct epoch count is used across training restarts. This enhances robustness and flexibility of training workflows and reduces misconfiguration risk for users. The change was implemented in commit 5715627ac0eb7f27d106f711c0571da28e4a7ef8 with cross-team collaboration (Co-authored-by: ppegolo <ppegolo@sissa.it>; frostedoyster <bigi.f@libero.it>).

2025-12 Monthly Summary for metatensor/metatrain Key features delivered: - Unified Data Collate Function for Training and Validation: Refactored the Trainer collate logic to use a single collate function for both training and validation datasets, reducing code duplication and simplifying maintenance. Commit 620072ba15f9274d58c5c76c0377308a1bae90b3. - Gradient Clipping Documentation Clarification: Updated docstrings and user docs to clarify the default gradient clipping value and the grad_clip_norm parameter, improving user understanding and reducing misconfigurations. Commit abf578cf39285a8125ba861742d155407e42ed7b. Major bugs fixed: - Removed redundant validation collate function (#978): Eliminated duplicate validation collate function to streamline data flow and reduce confusion. Commit 620072ba15f9274d58c5c76c0377308a1bae90b3. Overall impact and accomplishments: - Improved maintainability of the training data pipeline by eliminating redundancy and clarifying configuration requirements, enabling faster onboarding and fewer defects in future experiments. - Enhanced user guidance around gradient clipping, reducing support questions and misconfigurations. - Leaner, clearer codebase with focused responsibilities in data collation across training/validation. Technologies/skills demonstrated: - Python refactoring, documentation enhancement, and commit-level traceability. - Clear PR/message hygiene and collaboration signals through concise commit messages.

November 2025: Implemented TensorMaps contiguity enhancement for DiskDataset I/O in metatensor/metatrain, improving data integrity and disk I/O efficiency for large datasets. The change ensures TensorMaps are contiguous before writes, reducing fragmentation and enabling faster persistence. Commits: 2475bef0dc1069e5509796378391f9ce3593cf6a. Impact: more reliable data writes, smoother downstream ML workloads; alignment with performance and reliability goals.

For Oct 2025, metatensor/metatrain delivered a set of architecture migrations and code-cleanups that reduce technical debt, enhance training/inference accuracy, and enable future feature work. Highlights include migration to CompositionModel, removal of deprecated loss implementations, a granular Scaler scaling approach, explicit augmentation parametrization, and a corrected evaluation masking flow with improved device/dtype handling. These changes improve maintainability, reduce risk, and strengthen evaluation correctness across the pipeline.

September 2025 performance summary: delivered focused documentation improvements across metatensor/metatrain and lab-cosmo/pet-mad, enhancing accuracy, consistency, and user guidance. All changes were documentation-only, aligning terminology with current capabilities and reducing potential confusion for users and downstream teams.

Month: 2025-08. For repository metatensor/metatrain, delivered core CompositionModel improvements and spherical representations support, along with checkpoint upgrade safety across models. The work enhanced distributed training reliability, streamlined integration with SOAP-BPNN, and expanded testing for spherical representations, contributing to system stability and scalable experimentation. Business value includes safer upgrade paths, improved data handling, and faster iteration cycles on advanced representations.

July 2025 performance summary for metatensor/metatrain. The team delivered robust feature and data handling improvements, stabilized training workflows, and enhanced documentation, resulting in clearer developer feedback loops and improved model throughput. Changes focused on device correctness, data pipeline reliability, and centralized batching to reduce runtime errors and enable richer experiments.

June 2025 for metatensor/metatrain: Delivered a major architectural refactor of the CompositionModel to support a deprecation path, updating PET and related components to the new structure. This improves maintainability, reduces future debt, and enables smoother migrations. No major bugs fixed in this period; the focus was on architecture and code organization with a minimal external impact.

May 2025 monthly summary for metatensor/metatrain. Focused on delivering data handling improvements, configurability, and observability to support robust training workflows. No major bugs fixed this month; progress centered on feature delivery, validation, and documentation to improve reliability and repeatability.

April 2025 – Metatensor/metatensor: Focused improvements in data filtering reliability and practical guidance for neural network workflows. Delivered targeted robustness enhancements to the key filtering path and introduced practical PyTorch-based NN examples, enabling faster onboarding and more dependable model-building with metatensor. Key deliverables: - Bug fix: Key Filtering Robustness and Efficiency Improvement. Refactored filter_blocks to iterate directly over to_keep_indices, preserving key order and adding test assertions to ensure order remains correct post-filtering. Commit cf27a9cb8ec0ef5858b4645fda96c02731463e21. Impact: improved runtime efficiency, correctness, and test coverage for key filtering (#885). - Feature: Metatensor Torch Learn NN Examples and Usage Guides. Added example files demonstrating usage of metatensor.torch.learn.nn modules (Linear, ReLU, Sequential) and advanced constructs (EquivariantLinear, InvariantReLU, ModuleMap) to guide practical model construction and training. Commit 77a6bb41e7c997424591d968930d27624b192318. Overall impact and accomplishments: - Increased reliability and performance in a core data filtering path, reducing downstream data-related errors and latency in common workflows. - Expanded user education and adoption pathways through concrete PyTorch-based examples and usage guides, accelerating model development with metatensor. - Strengthened test coverage around filtering behavior and key ordering to support long-term maintainability and confidence in future changes. Technologies/skills demonstrated: - Python refactoring and benchmarks for performance-critical paths - Test-driven development and expanded test assertions for data order guarantees - PyTorch integration and practical NN module usage with metatensor.torch.learn.nn - Documentation and onboarding content creation for developer and user communities

March 2025 (2025-03) for metatensor/metatrain focused on robustness and clarity in distributed training workflows. No new features released this month; two critical bugs resolved that improve developer experience and documentation quality. These fixes reduce misconfiguration risks in multi-GPU distributed training and correct documentation for a common fitting workflow.

February 2025: Delivered a new Labels::difference API in metatensor/metatensor enabling set-difference computation between two Labels objects, with optional position mapping and robust error handling for mismatched label names. Backed by comprehensive unit tests, this supports data reconciliation and diff reporting in label-driven workflows, improving data integrity and traceability across production pipelines.

In November 2024, the Metatensor project focused on correctness and reliability in the Metatensor-Learn neural network modules. Delivered a critical bug fix for invariant_keys initialization and strengthened test coverage to guard against regressions. Key change: fixed default invariant_keys shape from (-1, 1) to (-1, 2) in metatensor-learn modules and added a regression test (test_default_invariant_keys) in tests/linear.py to validate the correction. The fix is tracked under commit 1fcc5240e47e55d0816e79279180bbea54b4df4f (Fix bug in default value for `invariant_keys` in `learn.nn` modules (#785)). Impact: ensures correct initialization of invariant keys, reduces risk of downstream errors in learning models, and improves overall reliability of the Metatensor-Learn stack. Technologies/skills demonstrated: Python, pytest, test-driven development, neural network module correctness, repository maintenance.