* feat: Add optional curriculum support to dataset registration and creation
* docs: Add docstrings to create_curriculum() and register_dataset()
* feat: Add curriculum configuration classes for CurriculumExperiment
* feat: Add weight parameter to CurriculumAttributeConfig and use in DatasetSpec
* refactor: Simplify CurriculumAttributeConfig with "*" attribute level support
* test: Add unit tests for CurriculumExperiment class
* feat: Add from_yaml() method to CurriculumExperimentConfig with unit test
This is a variant of the Game of Life task, which rather than trying to test the algorithmic simulation, tests the ability of the model to do explanatory reasoning of the board. The idea is that a model with good explanatory reasoning will be able to see that a game will not halt without simulating it into the future.
The task presents a GoL board, and the model is asked to predict if the board will halt (die, all cells zero) after n steps. Sometimes, the board will be made up of 'oscillators', isolated structures which never die. Othertimes, it is filled with non-oscillators, structures which will always die after a few steps. The model should deduce which case the presented board is.
* notebook for prepping first set of raw code files
* updated codeio processing notebook for repo-level processing
* fix for edge case in codeio scoring
* Add reformat notebook
* filtering pass
* add non-determinism filtering
* Tweak CodeIODataset & include first real data
* add basic codeio test, metadata
* math prompt improvements
* ignore brackets in complex_arithmetic results
* improve additional instruction in prompt of polynomial_equations
* more strict tests for score_answer in polynomial_equations
* simplify special reward handling
* fix test_intermediate_integration
* fix sokoban dataset
* add common dataset score_answer consistency test
* remove strip from ProceduralDataset::core score_answer(), strip in extract answer (optional, default=True)
* test: Move test_extract_answer() from test_dataset.py to test_utils.py
* refactor: Improve decimal reward computation with more flexible comparison
* fix: Implement rounding for format_number when round_if_needed is True
* test: Add test case for compute_decimal_reward with sign and zeros
