|
| 1 | +# Changelog |
| 2 | + |
| 3 | +All notable changes to the ASSTF project will be documented in this file. |
| 4 | + |
| 5 | +The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). |
| 6 | + |
| 7 | +## [1.0.0] - 2026-07-05 |
| 8 | + |
| 9 | +### Added |
| 10 | + |
| 11 | +- Initial public release of the ASSTF reference implementation. |
| 12 | +- `ASSTFLinear` and `ASSTFConv1d` drop-in PyTorch layers. |
| 13 | +- `BilevelTrainer` for alternating core/structural parameter updates. |
| 14 | +- `SurpriseMinimizer` for inference-time structural adaptation. |
| 15 | +- Six reproducible application demos (gesture, wake-word, anomaly, few-shot, RL, edge NLP). |
| 16 | +- Open data-generation scripts for all synthetic demos. |
| 17 | +- Real SST-2 validation for the edge NLP demo. |
| 18 | +- English-first documentation: Introduction, Architecture, Usage, API, and Open-Source Stages. |
| 19 | +- Dual licensing: Community License for non-commercial use, Commercial License for commercial use. |
| 20 | +- `pyproject.toml` and PyPI package `leanai-asstf`. |
| 21 | +- GitHub Actions CI workflow. |
| 22 | +- Community files: `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, `SECURITY.md`, `CHANGELOG.md`. |
| 23 | + |
| 24 | +### Fixed |
| 25 | + |
| 26 | +- `structural_gate` initialized to a negative value for near-zero structural contribution at startup. |
| 27 | +- `SurpriseMinimizer` no longer forces `model.train()` during adaptation. |
| 28 | +- Self-supervised adaptation loss now produces meaningful gradients. |
| 29 | +- App 05 RL adaptation target changed from self-distillation to policy-gradient objective. |
| 30 | +- `BilevelTrainer.evaluate` now correctly returns custom metrics. |
| 31 | +- `ASSTFConv1d` now honors the selected `svd_method`. |
| 32 | +- `shared.metrics.accuracy` broadcasting edge case resolved. |
| 33 | + |
| 34 | +--- |
| 35 | + |
| 36 | +## Unreleased |
| 37 | + |
| 38 | +### Planned |
| 39 | + |
| 40 | +- Real-world benchmarks on public datasets for Apps 01–05. |
| 41 | +- Hugging Face `peft` integration. |
| 42 | +- Edge deployment examples (Raspberry Pi, Arduino, ONNX). |
| 43 | +- arXiv technical report. |
| 44 | +- Community benchmark leaderboard. |
0 commit comments