v1.8.9

[1.8.9] - 2024-11-25

Changed

• Update launch and eval scripts for the eval phase of the contest
• Further improve alternative baseline agents
• Improve a few baseline agents

Added

• Add another time and place module for reality show
• Add support for scene-wise computation of metrics
• Add alternative versions of basic and rational agent factories

Fixed

• Catch another type of together API exception
• Fix time serialization in associative_memory

v1.8.8

[1.8.8] - 2024-11-13

Fixed

• Fixed a bug in the Schelling diagram payoffs component preventing some players
  from seeing outcomes on certain steps.

v1.8.7

[1.8.7] - 2024-11-5

Added

• add parochial universalization agent to the main factories
• add Jinja2 to the requirements to improve the prompting experience
• Add get_raw_memories and get_raw_memories_as_text functions on the memory component.

v1.8.6

[1.8.6] - 2024-10-29

Added

• Add ability to save and load rational and basic agents to/from json.
• Add a version of agent development colab that uses GCP hosted model

v1.8.5

Fixed

• Fix together.ai wrapper choice sampling procedure

v1.8.4: Increment version number to 1.8.4

[1.8.3] - 2024-10-19

Changed

• Allow conversation game master to be configured with memory and additional components
• Prevent daily_activities component from adding irrelevant info to memory and improve some of the time_and_place setting prompts.
• Add deprecation warning to the Sequential component
• together and mistral language model wrappers to no longer delete after last ".".
• Improve together.ai language model wrapper.
• disable pytype on a kwargs unpack
• Forward device and api_key only when explicitly specified.

Added

• Add support for custom cloud models (google cloud custom models)
• added instructions in the comment for finding relevant endpoint info. Defaulting location to "us-central1".
• Add agreement tracker contrib component.
• Add state_formation environment
• Add basic_agent_without_plan and use it in a scenario for state_formation.

Fixed

• Fixes a bug in inventory and schelling payoffs components, which previously crashed if player_action_attempt had ': ' in its text.

v1.8.3

[1.8.3] - 2024-10-08

Changed

• Jitter sleep times in together ai wrapper. This may improve speed.
• Change AllSimilarMemories component so that it no longer weights retrieval by recency or importance. This will make it function more like its name suggests it should. Note that many agents use this component so their behavior may change as a result of this change.
• make default observations summary time frame be 24h
• Add seed parameter to all simulations, propagating it into sample_parameters

Fixed

• Fix together AI api

v1.8.2

[1.8.2] - 2024-10-01

Changed

• Improves logging of strings with newlines

Fixed

• Insure concurrent execution of zero tasks returns immediately.
• Multi-item haggling environment, where players pick not only the price, but also the item to trade.

v1.8.1: Increment version number to 1.8.1

[1.8.1] - 2024-09-20

Added

• A wrapper that allows to limit the number of calls made by an LLM before it turns into a noop model

Fixed

• Make Inventory component thread safe.

v1.8.0

[1.8.0] - 2024-09-17

Changed

• Replace deprecated reference to langchain.llms with langchain_community.llms
• Improve logging in game master trigger components.
• Correct docstrings in base component type.
• Improve info printed on error in the together_ai model wrapper.
• Rename contrib agent components: contrib.components.agent.v2 -> contrib.components.agent.
• deprecate older contrib agent components
• Reformat reporting of subcomponents in all components derived from question of recent memory. We previously prefixed all subcomponents with "{name}'s\n" since typical names were things like "hunger" or "self identity". But it no longer makes sense given the way usage has evolved. Now we typically use subcomponents with keys like "Question: Is Bob hungry?\nAnswer: ". So the prefix should change.
• Make it optional to include full episode summaries in HTML logs and turn it off for the contest environments. The default preserves the old behavior.
• Improve Gemma2 wrapper: avoid lists, especially "\n\n" delimited lists.
• Improve clarity of GameMaster and fix type annotations
• Fix a few more issues in together.ai wrapper and improve filename handling in eval script.
• remove unused concurrent_action functionality
• Delete outdated base class and hide private methods
• Improve error handling, mid-sentence cutoffs, and max_tokens limit in Gemma2 model wrapper.
• Make formative memories warning less scary to participants.
• Print less warnings from mistral wrapper when they do not indicate a real problem
• add together_ai to utils
• Use concurrency helpers and hide executor implementation
• Add a maximum number of steps to the conversation scene. Default is 3
• Improve reality_show action/reward observations, parameterize more of its settings, and add scenarios.
• remove unused dependency
• Print recoverable exceptions in mistral model wrapper.

Added

• Bots now support more complex fixed rule policies.
• Add an option to prevent the inventory component from ever increasing amounts of items in inventories (off by default, preserving the old behavior). This change also adds default context for the inventory reasoning chain of thought.
• Add an agent with a very simple prompting strategy, just observe and recall relevant memories. Nothing else.
• Add paranoid supporting agent and a substrate that uses it.
• Add test for supporting agent factories and add paranoid agent to the main role factories test. Also give basic_puppet_agent a default fixed_response (no fixed responses) for testing with no parameters.
• Add Together AI language model to Concordia. This can serve Gemma 2 9B
• Add additional concurrency helpers run_tasks and run_tasks_in_background
• allow colab users to pass an api key explicitly.

Fixed

• Avoid pylint error by using direct import
• Fix pytype error
• Add lock to prevent duplicated _make_pre_act_value calls
• Hold lock on memory read to prevent reading during a write.
• Fix a bug which can arise when a model does not precisely follow the requested format in AccountForAgencyOfOthers.
• Add locking and check phase transitions
• The previous change didn't fully suppress harmless mistral exceptions, this one fixes it.
• Ensure run_parallel uses >1 worker
• Fix silent failures by checking futures after completion
• Fix language model setup util