v0.17.5: DOI generation

✨ New features and improvements

• This is a dummy release to test whether DOI generation via Zenodo works.

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.17.4...v0.17.5

v0.17.4: fix: Fix for fix of eager engine access during init

✨ New features and improvements

-

🔴 Bug fixes

• https://github.com/MantisAI/sieves/releases/tag/v0.17.3 had an issue preventing it from properly fixing the eager access issue. This is properly patched here. (#182 )

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.17.3...v0.17.4

v0.17.3: Fix eager engine access during init

✨ New features and improvements

-

🔴 Bug fixes

• Engine classes were accessed eagerly during engine intialization, even without the corresponding library installed. This is fixed in this patch - engine classes are only accessed during init if the corresponding library is installed. (#181)

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.17.2...v0.17.3

v0.17.2: `transformers[torch]` as default dependency; tests run against built wheel

✨ New features and improvements

• transformers[torch] is now a default dependency. (#180)

🔴 Bug fixes

• CI tests are now run against a built version as to avoid mismatches between library behavior when run from source vs. built. (#180)

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.17.1...v0.17.2

v0.17.1: Fix ingestion import with uninstalled requirements; drop support for `unstructured`

✨ New features and improvements

• After recent refactorings importing ingestion tasks failed with ingestion libraries (docling, marker) not installed. This is fixed by this patch. (#179)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

• unstructured is no longer supported for ingestion. (#179)

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.17.0...v0.17.1

v0.17.0: Task optimization; drop support redundant generation engines

✨ New features and improvements

• Task optimization: optimize prompts and few-shot example selection with task.optimize(). All tasks are supported. (#178)
• Upgrade chonkie support from v0.x to v1.x (#177)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

• Explicit support for the following structured generation tools has been cut, as they either can be used via one of the other explicitly supported tools or didn't add relevant features (#178):
  • ollama - you can still use sieves with ollama, just wrap your ollama model instance in a dspy or langchain model.
  • vllm - some reasoning: you can still use vllm, just wrap the corresponding model object in a dspy or langchain model.
  • instructor - didn't add relevant features that weren't added by any of the other, still supported tools.

📖 Documentation and examples

-

👥 Contributors

@rmitsch, @ivyleavedtoadflax

---

Full Changelog: v0.16.1...v0.17.0

v0.16.0: Fix `yield` flow; drop `Distillation` task in favor of task-wise distillation

✨ New features and improvements

-

🔴 Bug fixes

• Fix broken yield to allow for true doc-/batch-wise processing (#172)

⚠️ Backwards incompatibilities

• Distillation is available only via task interface (i.e. task.distill(...)), no more dedicated Distillation task. This was done as the Distillation task necessarily requires processing the entire dataset at once, the nature of the task doesn't allow for iterative processing - this made it fundamentally incompatible with the iterative processing style all other components supported. (#173)

📖 Documentation and examples

-

👥 Contributors

@rmitsch, @ivyleavedtoadflax

---

Full Changelog: v0.15.1...v0.16.0

v0.15.1: Async batching for Outlines models; remove model default values

✨ New features and improvements

• Use async batching for models with Outlines when Outlines doesn't support batching natively (#170)

🔴 Bug fixes

• Remove default model values like max_tokens that could cause crashes for model not supporting these kwargs (#169)

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch, @ivyleavedtoadflax

---

Full Changelog: v0.15.0...v0.15.1

v0.15.0: Slimmer default depency group; fix distillation bug

✨ New features and improvements

-

🔴 Bug fixes

• Fix classification distillation (#164)

⚠️ Backwards incompatibilities

• Ingestion-related libraries are now in the optional ingestion dependency block. Make sure you install all required dependencies. (#163)

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.14.0...v0.15.0

v0.14.0: Simpler API via dropping `Engine`

✨ New features and improvements

• Phase out Engine as user-facing concept, accept models directly instead (#161)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

• Engine is no longer accepted as argument for tasks. Pass the model instead. (#161)

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.13.0...v0.14.0