feat: training set SDK support (DE-8692)

CHANGELOG.md

@@ -5,6 +5,27 @@ All notable changes to the [Nucleus Python Client](https://github.com/scaleapi/n
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.21.5](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.21.5) - 2026-08-31
+
+### Changed
+- **Training set membership scales to whole datasets/slices (DE-8692).** Adding or removing a whole dataset, slice, or the members of a large training set now streams asynchronously server-side instead of being resolved in memory, so a training set can span an arbitrarily large source. Small, explicit `item_ids` / `items` changes stay synchronous. When a change kicks off an async build the response carries a `job_id`; `create_training_set()`, `add_training_set_items()`, `remove_training_set_items()`, and `create_training_set_version()` block on it by default (`wait_for_completion=True`). `TrainingSet.status` reads `"building"` while a job is streaming members and `"ready"` otherwise.
+- **`remove_training_set_items()` / `TrainingSet.remove_items()` accept sources.** Removal now takes the same source shape as add — explicit `item_ids` / `items`, or a whole `slice_id(s)` / `dataset_id(s)` / `training_set_ids` whose (potentially huge) membership is streamed out — with `wait_for_completion` / `verbose`. `item_ids` stays the first positional argument, so existing `remove_items([...])` / `remove_training_set_items(id, [...])` calls are unchanged.
+
+### Removed
+- **`scene_ids` on training set add/create.** Training sets resolve dataset items only (there is no scene source, unlike benchmarks); the parameter never had a server-side effect and has been dropped from `add_training_set_items()` / `TrainingSet.add_items()`.
+
+## [0.21.4](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.21.4) - 2026-08-27
+
+### Added
+- **Training set export / download (DE-8692).** Pull a training set's members back out as fully-hydrated items. `TrainingSet.export_items()` / `NucleusClient.export_training_set_items()` page the whole set and return `DatasetItem`s (media location, `reference_id`, `metadata`, `width` / `height`, and the server-side `dataset_item_id`). `TrainingSet.export_to_file(path)` writes every member to a JSONL file (one raw export record per line — `dataset_item_id`, `dataset_id`, `reference_id`, `metadata`, `image_location`, `pointcloud_location`, `width`, `height`) and returns the count written. `TrainingSet.download_items(directory)` streams each member's media file to disk (named by `reference_id`, falling back to `dataset_item_id`), returning the number downloaded.
+
+## [0.21.3](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.21.3) - 2026-08-27
+
+### Added
+- **Training sets (DE-8692).** `TrainingSet` — a mutable, versioned, model-scoped collection of `dataset_item` ids spanning one or more datasets. Create and attach one to a model with `NucleusClient.create_training_set(name, model=...)` or `Model.create_training_set(name, ...)`, providing members through any combination of `item_ids`, `(dataset_id, reference_id)` pairs via `items`, `slice_id` / `slice_ids`, `dataset_id` / `dataset_ids`, and the members of other training sets via `training_set_ids`. Fetch/list with `get_training_set()` / `list_training_sets()`; read a model's pinned set via `Model.training_set`.
+- **Mutable membership.** Add sources with `TrainingSet.add_items()` / `NucleusClient.add_training_set_items()` (async, same sources as create), remove with `TrainingSet.remove_items()` / `NucleusClient.remove_training_set_items()`, and page members with `TrainingSet.items()` / `NucleusClient.list_training_set_items(limit=, offset=)`.
+- **Versioning / lineage.** Cut a new version with `TrainingSet.new_version()` / `NucleusClient.create_training_set_version()` (child inherits the parent's items, sources add on top, `removed_item_ids` prune; `parent ∪ added ∖ removed`), or pass `parent_training_set_id` to `create_training_set()`. Version defaults to a minor bump; pass `bump_type="major"` or explicit `version_major` + `version_minor`. Inspect a set's lineage with `NucleusClient.list_training_set_family()` and repin a model to a specific version with `Model.repin_training_set()` / `NucleusClient.repin_training_set()`. `TrainingSet` exposes `model_id`, `parent_training_set_id`, `version_major`, `version_minor`, and `version_label`.
+
 ## [0.21.2](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.21.2) - 2026-08-17
 
 ### Added