SDK-43: Add LLM unlearning support

[benglewis]

Summary

• add SDK models for LLM resources covering Hugging Face sources, unlearning datasets, and runs
• expose the new LLM helpers from the package root for easier imports
• exercise the new logic with sanity tests that validate request payloads and launch flow

Testing

• PYENV_VERSION=3.11.12 pytest tests/llm -q
• PYENV_VERSION=3.11.12 ruff check .

feature: SDK-43

---

Codex Task

---

Note

Adds a new hirundo.llm module to manage LLM resources and unlearning.

• Introduces BiasType/BiasRunInfo for configuring unlearning runs
• Adds HuggingFaceTransformersModel and LlmModel with create/list/get/delete and server payload serialization
• Adds UnlearningExample and LlmUnlearningData with create/list/get/delete for unlearning datasets
• Adds LlmUnlearningRun with launch and get for run lifecycle
• Exposes all new LLM helpers via package root __init__
• New tests validate request URLs/payloads and run launch flow

^{Written by Cursor Bugbot for commit c44cf64. Configure here.}

[benglewis]

This doesn't look right... Lots of things are off

[cursor]

Comment @cursor review or bugbot run to trigger another review on this PR

[cursor]

Truthiness check ignores zero organization ID value

Low Severity

The list_models and list_datasets methods use a truthiness check (if organization_id) instead of an explicit None check. This causes organization_id=0 to be treated as falsy and silently ignored. The type annotation int | None allows zero, and the create() methods in the same file correctly use if organization_id is not None:. This inconsistency could lead to unexpected behavior when explicitly filtering by organization ID zero.

Additional Locations (1)

• hirundo/llm.py#L204-L205