Skip to content

Latest commit

 

History

History
116 lines (99 loc) · 7.46 KB

CHANGELOG.md

File metadata and controls

116 lines (99 loc) · 7.46 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog. Given its evolving nature, the project does not yet adhere to Semantic Versioning.

[Unreleased]

  • feat: Add seed setting to GenConf. Commented-out because of lack of support in OpenAI models and some llama.cpp hardware accelerations. To be enabled when wider support becomes available.

[0.4.4]

  • feat: Support vision models from OpenAI, Anthropic and Llava-based local models.
  • feat: Add Msg class for better handling of other data types besides text, for now only images.
  • feat: Update Thread class to support images. Also added trimming functionality and as a result removed emptied Context class.
  • feat: Add close() method to Model* classes, to release resources.
  • fix: Removed no longer used _workaround1 in LlamaCppTokenizer.
  • fix: Avoid setting "max_tokens" in remote models that support it, for GenConf.max_tokens=0.
  • fix: Update configs to new OpenAI models.
  • docs: Add vision model and Thread use documentation.
  • docs: Add receipt image extraction example.
  • test: Add tests for Thread, Msg and vision models.

[0.4.3]

  • feat: Add support for dataclass Optional and Union annotations.
  • feat: Add Groq remote model provider.
  • fix: Add deepcopy to json_schema_massage() to solve schema massaging error.
  • fix: Set Thread.repr output the same as str for development convenience.
  • docs: Improve Pydantic and dataclass documentation with examples of Optional, Union and default fields.
  • test: Add tests for complex extraction into Pydantic and dataclass objects.

[0.4.2]

  • feat: Add Model.create() argument to retrieve the actual initialization params used to create the model.
  • fix: Correct OpenAI's "max_tokens_limit" setting to 4096 in base_models.json, a more sensible default value for future models.
  • fix: Update Model.version() formats to be simpler and simplify comparison between versions.

[0.4.1]

  • feat: Add Anthropic provider.
  • feat: Add chat template formats for Llama3 and Phi-3 instruct models, StableLM-2, Command-R/Plus.
  • feat: Add output_fn_name property to Model, for changing the output function name in models that use a Tools/Functions API.
  • feat: Better JSON/Schema decoding errors.
  • fix: Don't use a string representation of the dataclass when its doc string is unset, during JSON Schema creation, to keep equivalence with Pydantic-based generation.
  • fix: Workaround for MistralModel, where the Mistral API misses api_key argument/env variable when run from pytest.
  • fix: Consolidate all Model class info as methods to avoid property/method() calling confusion.
  • docs: Update installation instructions and include info on new Anthropic provider.
  • test: Better parametrized tests for remote and local models.
  • test: Add tests for new provider.

[0.4.0]

  • feat: New providers: Mistral AI, Together.ai and Fireworks AI allowing access to all their chat-based models.
  • feat: Model classes now support async calls with the '_async' prefix, for example extract_async(). This requires model API support: only remote models will benefit. Local models (via llama.cpp) can still be called with _async methods but do not have async IO that can run concurrently.
  • feat: Add 'special' field to GenConf, allowing provider or model specific generation arguments.
  • feat: All models now also accept model path/name starting with their provider names as in Models.create().
  • feat: Change Model.json() to stop requiring a JSON Schema as first argument.
  • fix: More robust JSON extraction for misbehaved remote models.
  • fix: LlamaCppModel no longer outputting debug info when created in Jupyter notebook environment with verbose=False.
  • fix: Default "gpt-4" model in 'sibila/res/base:models.json' now points to gpt-4-1106-preview, the first GPT-4 model that accepts json-object output.
  • docs: Add API references for new classes and _async() methods.
  • docs: Add new async example.
  • test: Add new tests for new providers/model classes.

[0.3.6]

  • feat: Migrate hardcoded OpenAI model entries from OpenAIModel to 'res/base_models.json'.
  • feat: OpenAI now accepts unknown models using defaults from 'openai:_default' key in 'res/base_models.json'.
  • feat: Support OpenAI models with a limit on max_tokens output values, like "gpt-4-turbo-preview" (input ctx_len of 128k but only up to 4k output tokens).
  • feat: Auto-discover maximum ctx_len in LlamaCppModel loaded files, when 0 is passed.
  • feat: Add negative int factor mode to GenConf.max_tokens setting, allowing for a percentage of model's context length.
  • fix: Add coherent error exceptions when loading local and remote models.
  • fix: Correct interact() error when GenConf.max_tokens=0.
  • fix: Correct several chat template formats.
  • test: Add many new tests for gpt-3.5/4 and llama.cpp models.
  • docs: Update tips section.

[0.3.5]

  • feat: Split Models factory config in two levels: base definitions in sibila/res and Models.setup() loaded definitions from user folders. These levels never mix, but a fusion of the two is used for models/formats resolution. Only in this manner can "models" folder definitions be kept clean.
  • fix: Option sibila formats -u is removed as result of the two-level Models factory.
  • fix: Correct delete of link entries in models.json and formats.json, which was resolving to targets (and deleting them).
  • fix: Raise ValueError when trying to generate from an empty prompt in LLamaCppModel.
  • fix: Update Models to check linked entries when deleting.
  • fix: Update template format discovery to work in more edge cases.
  • test: Add test cases for sibila CLI and LlamaCppModel.

[0.3.4]

  • feat: Improve template format discovery by looking in same folder for models/formats.json.
  • fix: Update legacy importlib_resources reference.
  • docs: Improve text.

[0.3.3]

  • fix: Move base_models.json and base_formats.json to sibila/res folder.
  • fix: Add base_models.json and base_formats.json to project build.
  • fix: Correct .gitignore skipping valid files.
  • docs: Update installation help and mentions to base_models/formats.json.

[0.3.2]

  • feat: Added sibila CLI for models and formats management.
  • feat: Added methods in Models class for CLI functionality.
  • fix: Blacklisting character control set in JSON strings grammar.
  • docs: Improved docs and added section about sibila CLI.
  • docs: Added CLI example.

[0.3.1]

  • feat: Improved documentation.
  • feat: Model.known_models() returns a list of fixed known models or None if unlimited.
  • feat: LlamaCppModel now also looks for the chat template format in a 'formats.json' file in the same folder as the model file.
  • feat: Added GenConf.from_dict() and renamed asdict() to as_dict().
  • fix: Creating a model entry in "models.json" with a genconf key was not being passed on model creation.

[0.3.0]

  • feat: Added Models singleton class that centralizes ModelDir and FormatDir.
  • feat: New extract() and classify() methods for type-independent extraction and classification.
  • feat: Renamed confusing gen() and gen_() method names to simpler alternatives type() and gen_type().
  • feat: Replaced dictype definitions with dataclasses, a better to extract dictionaries.
  • feat: Added version() and provider_version() to Model and children classes.
  • fix: Using 2 * "\n" to separate message text from automatically added json_format_instructors ("Output JSON", etc.), to provide more meaningful separation.
  • fix: Added requirement for package typing_extensions because of Self type and Python 3.9+ compatibility.