Add Tags to Samples from Generator

[jkeskingvillage]

A set of tags to describe what is in each data sample. Currently, this is what I am using:

avoid_hash_conflicts = {bool} False
constraints_in_goal = {bool} False
constraints_in_input = {bool} False
constraints_in_memory = {bool} False
constraints_in_output = {bool} False
constraints_in_spec = {bool} False
coupling_of_agents = {float} 75.0
enable_mapping_cost = {bool} False
enable_maps = {bool} True
enable_slots = {bool} True
enable_slotting_cost = {bool} False
enable_typing = {bool} False
failure_in_history = {bool} False
flat_type_hierarchy = {bool} False
history_in_memory = {bool} False
multiple_goals = {bool} True
number_of_agents = {int} 50
number_of_samples = {int} 1000
number_of_types = {int} 10
number_of_variables = {int} 50
objects_as_goal = {bool} False
objects_in_memory = {bool} False
operators_as_goal = {bool} True
or_goals = {bool} False
parameterized = {bool} False
parameters_per_agent = {int} 5
tristate_variables = {bool} False
upload = {NoneType} None

For example, if a sample has more than one goal, then multiple_goals should be True. The names should be self explanatory but ask me if you are confused. Here is the schema: link. Note that a request to generate data also has the same tags but also include how many to generate, while the tags for the data also include the final length of the plan. The length_of_sequence of the sequence is a tag that is not like the others because it is not used in the generator but its value is determined and recorded after the sample is generated.

Currently, we only need to cover the following tags, the rest keep the default.

number_of_agents
number_of_variables
parameters_per_agent
coupling_of_agents
enable_slots
enable_maps
multiple_goals
objects_in_memory
length_of_sequence

In terms of code, the only change required here is at the time of generating a sample, or after it is done, add this set of tags as a field "tags" in the data, inside the agent_info_generator_input field.

• In the training samples, it seems this appears twice. Once inside the agent_info_generator_input and again inside agent_info_generator_input inside agent_info_generator_output_item. I don't need both but any one or both is fine.

• Similarly, in the testing samples (in the final report), it appears twice: inside pddl_generator_output inside llm_response_planning_data. Again both, or any one, will do.

I have added some examples of the train (produced from the profiler) and test files (produced in the final report after validation in your pipeline) here. train_samples.json is a sample training file; and testing_llama.json / testing_granite.json are sample test files.

[jkeskingvillage]

@TathagataChakraborti Can an instance of this tags object be return from a method in a Flow instance? Profiler can pass the tags instances to the downstream. The fields in the object seem to be handled best at Flow. I can handle the downstream once Flow has a method to return an instance of this Tags object.

[TathagataChakraborti]

Being addressed in iss104 branch ETA Wednesday

[jkeskingvillage]

This issue will be handles at the Runner.

[jkeskingvillage]

This can be handled best at PddlGeneratorOutput. @TathagataChakraborti I will add a method to produce tags at PddlGeneratorOutput to pass the tags to the downstream.

[jkeskingvillage]

Done.