Skip to content

v1.3.0
Compare
Choose a tag to compare
@jzleibo jzleibo released this 04 May 18:12
v1.3.0
01fcf89

Changed

  • Improve initialization of self perception component.
  • Add configurable call_to_speech in conversation scene
  • Change observation components to retrieve only observations from memory, filtering away memories produced by other components.
  • Goal achievement and morality metrics trigger on action attempt rather than observation.
  • Prevent the game master from inventing direct quotes
  • Use new unstable attribute in pyink settings
  • Add state lock and only update if the memory has changed.
  • Replace tensorflow ST5 embedder with torch based one
  • Support passing axis into plotting functions
  • Fixing the call sequence in the apply_recursively and adding concurrency.
  • Minor fixes to enable support of pandas 2.1
  • Added len to associative memory, which returns the number of entries in the memory bank.
  • Add clock to the plan component. It makes more sense when re-planning to know the time. Also, enables doing only one update per step
  • Make sure components update only once per time step. Now that updates are recursively called on sub-components this makes a significant saving in API calls.
  • Improve formative memories factory. Make more plausible memories.
  • Make action_spec configurable per step
  • Improve precision in the Inventory component.
  • GM step function can optionally specify specific players to act this round.
  • Skip updating dialectical reflection if already done on current timestep.
  • More verbose uncertainty scale questions
  • Improve error message for out of range timestamps (<1677 CE or >2262 CE).
  • Encapsulate default game master instructions inside game_master class.
  • Prevent adding duplicate memories
  • Adding a check that makes sure the three questions components only updates once per time step
  • increase max num tokens that can be produced by the direct effect component.

Added

  • Add component that implements creative reflection
  • Add customisation options for the conversation scene
  • Add a current scene component to make the game master aware of the scene
  • Adding the recursive component update to the game master, same logic as in the agent.
  • Add update_after_event call to components in the agent, which is called with the action attempt as argument. This brings GM and agent flow of calls closer and enables components to update given the action attempt (e.g. reflect on action).
  • Add updated ollama models
  • Add scene runner and associated types
  • Add example result obtained from running the three_key_questions colab.
  • Add agent component to select all relevant memories
  • Add relationships component that tracks the status of relationships between agents.
  • Add dialectical reflection component. Agents with this component reflect on a specific topic and incorporate some academic-oriented memories like books they have read. Also add the scheduled hint component which makes it possible to send specific text to a specific player when specific conditions are met.
  • Add a function to create example question-answer pairs on an interactive document.
  • Add optional thought functions to preserve direct quotes of agent speech.

Fixed

  • Fix bug in 'all similar memories' component
  • Fix pytype error
  • Fix a bug in the scene runner.
  • Don't install tensorflow-text on M1 OSX.
Assets 2
Loading