feat: training set SDK support (DE-8692)

[luke-e-schaefer]

Training Set SDK support (DE-8692)

Adds SDK support for training sets — a mutable, versioned, model-scoped collection of dataset_item ids spanning one or more source datasets. Similar to the Benchmark resource.

Ticket: DE-8692

What's added

• nucleus/training_set.py — TrainingSet dataclass: from_json, refresh, update, delete, items, add_items, remove_items, new_version, family, lineage fields.
• NucleusClient methods: create_training_set, get_training_set, list_training_sets, get_model_training_set, repin_training_set, update/delete_training_set, list/add/remove_training_set_items, create_training_set_version, list_training_set_family.
• Model entry points: Model.create_training_set(...), Model.training_set, Model.repin_training_set(...).

Greptile Summary

The PR adds model-scoped training-set creation, mutation, versioning, export, and media-download support.

• Exposes training sets through NucleusClient, Model, and the top-level package.
• Adds payload and response DTOs for membership, lineage, and asynchronous operations.
• Adds paginated JSONL export and per-dataset media downloads with collision handling.
• Adds unit coverage for the new SDK workflows.

Confidence Score: 4/5

The PR is not yet safe to merge because downloads can still silently replace members whose generated paths alias on case-insensitive filesystems.

Collision tracking compares exact path strings, but supported filesystems can resolve differently-cased strings to the same entry, allowing the later media transfer to overwrite the earlier one.

Files Needing Attention: nucleus/training_set.py

Important Files Changed

Filename	Overview
nucleus/training_set.py	Adds the training-set resource and local export/download helpers; collision detection can still overwrite case-aliased destinations on case-insensitive filesystems.
nucleus/init.py	Adds client APIs for training-set CRUD, membership, versioning, pagination, and export.
nucleus/model.py	Adds model-level convenience entry points for creating, reading, and repinning training sets.
nucleus/data_transfer_object/training_set.py	Adds the typed training-set membership-page response.
tests/test_training_sets.py	Covers the added workflows and exact-string download collisions but does not exercise case-insensitive destination aliases.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  User[SDK caller] --> Model[Model]
  Model --> TrainingSet[TrainingSet]
  TrainingSet --> Membership[Add or remove members]
  TrainingSet --> Version[Create version]
  TrainingSet --> Export[Page export records]
  Export --> JSONL[Write JSONL]
  Export --> Download[Download media by dataset]

Loading

Prompt To Fix All With AI

### Issue 1
nucleus/training_set.py:330
**Case-aliased downloads still overwrite**

If `download_items` runs on a case-insensitive filesystem, members whose generated paths differ only by case, such as `Frame.jpg` and `frame.jpg`, bypass the exact-string `used_paths` check. The second transfer then replaces the first filesystem entry while the method reports both members as downloaded.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (6): Last reviewed commit: "flexing" | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

Context used:

• Knowledge Base — Python client API
• Knowledge Base — Model workflows

[greptile-apps]

Case-aliased downloads still overwrite

If download_items runs on a case-insensitive filesystem, members whose generated paths differ only by case, such as Frame.jpg and frame.jpg, bypass the exact-string used_paths check. The second transfer then replaces the first filesystem entry while the method reports both members as downloaded.

Prompt To Fix With AI

This is a comment left during a code review.
Path: nucleus/training_set.py
Line: 330

Comment:
**Case-aliased downloads still overwrite**

If `download_items` runs on a case-insensitive filesystem, members whose generated paths differ only by case, such as `Frame.jpg` and `frame.jpg`, bypass the exact-string `used_paths` check. The second transfer then replaces the first filesystem entry while the method reports both members as downloaded.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.