📣 🥁 Changelog & Releases

📍 Unreleased Version X.X.X - XXXX-XX-XX

🛠 Fixed

  • Moved ruff config from pyproject.toml to ruff.toml; moved pytest options to pytest.ini.

  • Pointed CONTRIBUTING.md at ruff.toml / pytest.ini.

  • Ran poe sphinx in the tox env (docs build gate).

📍 Version 0.3.0 - 2026-07-15

🛠 Fixed

  • Auto-bump KTO per_device_train_batch_size to at least 2 (TRL requires batch size > 1 for the KL term).

  • Moved mypy config from pyproject.toml to mypy.ini; added tox.ini (poe tox).

  • Fixed tox to use tox-uv (uv-venv-lock-runner + only-managed) so each env uses uv-managed CPython from the env name (not the project .venv).

🔥 Added

  • Preference objectives: DPOObjective, ORPOObjective (TRL experimental), KTOObjective.

  • Preference / KTO JSONL loaders and demo datasets (about_amir.prefs.jsonl, about_amir.kto.jsonl).

  • CLI --objective sft|dpo|orpo|kto and --beta; examples run_dpo_lora.py / run_kto_lora.py + poe train-dpo / poe train-kto.

  • tox for multi-Python check/test/build (aligned with afk-bot).

  • Python 3.13 support (requires-python = ">=3.10,<3.14", CI + tox matrix).

📍 Version 0.2.0 - 2026-07-14

🛠 Fixed

  • Fixed AdaLoRA training callback to call update_and_allocate on on_optimizer_step (grads must exist before zero_grad).

  • Fixed AdaLoRA tiny-SFT defaults (warmup tinit / tfinal, schedule estimation) so probes can pass.

  • Fixed LLM judge score parsing (ignore 0-10 scale echoes) and digit-constrained scoring for small judge models.

  • Fixed AdaLoRA schedule dataclasses.replace typing for mypy.

  • Renamed CLI entry point from slick-tune to slicktune (matches the package name).

  • Hardened docs CD: deploy _static/ in a separate FTP sync so the host does not FIN mid-session on first publish.

🔥 Added

  • Strategies: DoRAStrategy and AdaLoRAStrategy (+ AdaLoRACallback).

  • Holdout eval stack: slicktune eval, perplexity, SubstringJudge / LLMJudge, and about_amir.eval.jsonl.

  • Example scripts for DoRA / AdaLoRA SFT runs.

  • Sphinx docs (Furo + AutoAPI + MyST), Fine-Tuning Visual Guide, and SemVer-style changelog.

  • Docs CD workflow: build Sphinx on push to master and deploy to docs.slickml.com/slick-tune/ via FTP.

  • Keyword-only public APIs and named-arguments Cursor rule.

  • Added slicktune --version (Click version_option).

📍 Version 0.1.0 - 2026-07-14

🔥 Added

  • Initial Phase 1 library: composable Tuner with model × strategy × objective × data × metrics.

  • Strategies: LoRAStrategy, QLoRAStrategy, FullStrategy.

  • Objective: SFTObjective (DPO stub for Phase 3).

  • Data loaders for SFT / probe JSONL and shipped about_amir train / probe datasets.

  • Metrics tracker, probe recipes, and slicktune train / probe CLI plus example scripts.

  • CI (uv, Ubuntu + macOS, Python 3.10–3.12), Codecov, 100% unit-test coverage gate.