Closed TathagataChakraborti closed 1 month ago
@TathagataChakraborti Circular constraints happen due to the nature of algorithm to create coupling. We can re-visit this if large number of actions is needed. Meanwhile, this is, also, not rare in the real world APIs (i.e. id
, name
, token
, time
, ...).
@TathagataChakraborti Let's leave double underscores in names as they are. This can be useful for parsing responses from LLMs. Everything else will be done.
Done.
Some more changes are needed.
Example planning problem descriptions. planning_examples.txt
[x] For the slot filler line, the second line should go after the third it sounds more natural coz it says what ask does first before saying its less preferred.
[x] The slot fillable line should come after the introduction of the ask operation. The current description says which things cannot be asked before introducing the ask operation.
[x] We should probably move the slot filler description to after the actions, next to or even after the map description.
[x] In the action description, add a line like action A has parameters x, y, z before ... To execute action A, variable x, y, z should be known.
[x] Change the mapping line from Values for Variable v1 can be used for Variable v0 to include the map operation. Something like Action map can determine the value of Variable v1 from Variable v0.
[x] Uncapitalize action, variable, etc. Currently some are capitalized, some are not.
[x] ~Remove double underscores?~
Current
Desired
Circular constraints
I noticed that many samples have this pattern: the parameters are cycled through in order and back again.
This might impact the performance by being biased to a particular class of problems where inputs and outputs are circular. Can we randomize this? Or are other places randomized so it doesn't matter?