[WIP] Initial implementation of choice-only models

[digicosmos86]

This PR introduces choice only models. The following needs to be changed in order for this to happen:

1. Modified Config class and HSSM to add support for choice-only property fields
2. Temporarily implemented a dummy simulator function to get around the lack of simulators in SSMS
3. Implemented a general softmax_inv_temperature likelihood function for arbitrary numbers of logits in the parameter
4. Implemented softmax_inv_temperature_2 and softmax_inv_temperature_3 model configs, and added these to SupportedModels
5. DataValidator Mixin does not need to keep internal states. The states are kept in HSSM class itself
6. Added unit tests for the new likelihood function and simple tests for choice only models (will need to update this)

[Copilot]

Pull request overview

This PR introduces initial support for choice-only models in HSSM by adding a softmax inverse-temperature likelihood and corresponding model configs, plus updates to HSSM/distribution utilities and new tests.

Changes:

• Add a new analytical likelihood (softmax_inv_temperature) and default configs for 2- and 3-logit variants.
• Extend HSSM and distribution utilities to handle choice-only response shapes and model-building paths.
• Add unit/slow tests covering new configs and basic sampling for choice-only models.

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
src/hssm/likelihoods/analytical.py	Adds softmax_inv_temperature likelihood implementation.
src/hssm/modelconfig/_softmax_inv_temperature_config.py	Provides shared default config factory for softmax inverse-temperature models.
src/hssm/modelconfig/softmax_inv_temperature_2_config.py	Adds default config entry point for the 2-logit model.
src/hssm/modelconfig/softmax_inv_temperature_3_config.py	Adds default config entry point for the 3-logit model.
src/hssm/_types.py	Registers new supported model names.
src/hssm/config.py	Adds Config.is_choice_only property derived from response fields.
src/hssm/hssm.py	Adds choice-only handling for response formatting, missing-data logic, and RV fallback behavior.
src/hssm/distribution_utils/dist.py	Adds is_choice_only plumbing for RV signature and distribution logp behavior.
src/hssm/data_validator.py	Changes DataValidatorMixin to rely on subclass state rather than its own constructor.
tests/test_modelconfig.py	Adds tests for the new softmax inverse-temperature configs.
tests/test_config.py	Adds assertions for the new is_choice_only behavior.
tests/test_hssm.py	Adds a test for choice-only + deadline behavior in HSSM.
tests/slow/test_choice_only.py	Adds slow tests for choice-only sampling paths and likelihood shape checks.
.gitignore	Adds ignores for assistant/tooling-related files.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

softmax_inv_temperature is documented/used as a log-likelihood, but it currently returns exp(log_prob_choices) (a probability). make_distribution expects the log-likelihood (logp), and will apply exp() again in the lapse mixture, which will yield incorrect values / potential overflow. Return log_prob_choices (and consider renaming variables/docs accordingly).

Suggested change

	return pt.exp(log_prob_choices)
	return log_prob_choices

[Copilot]

Docstring typo: “A scaler for the softmax temperature” should be “A scalar …”.

Suggested change

	A scaler for the softmax temperature (0, inf).
	A scalar for the softmax temperature (0, inf).

[Copilot]

Removing DataValidatorMixin.__init__ is a breaking change for existing direct instantiations (the repository’s tests/test_data_validator.py constructs DataValidatorMixin(...)). Either restore a minimal/backward-compatible __init__ (even if only for tests) or update the tests and any downstream usage to instantiate a concrete class that sets the required attributes.

Suggested change

	def __init__(
	self,
	data: pd.DataFrame | None = None,
	response: list[str] | None = None,
	choices: list[int] | None = None,
	n_choices: int | None = None,
	extra_fields: list[str] | None = None,
	deadline: bool = False,
	deadline_name: str | None = None,
	missing_data: bool = False,
	missing_data_value: float = -999.0,
	is_choice_only: bool = False,
	) -> None:
	"""Initialize DataValidatorMixin with optional arguments.
	This minimal initializer is provided for backward compatibility with
	direct instantiation (e.g., in tests). Subclasses are free to override
	this method and set these attributes themselves.
	"""
	self.data = data if data is not None else pd.DataFrame()
	self.response = response if response is not None else []
	self.choices = choices if choices is not None else []
	# If n_choices is not provided, infer from choices if available.
	if n_choices is not None:
	self.n_choices = n_choices
	else:
	self.n_choices = len(self.choices) if self.choices is not None else 0
	self.extra_fields = extra_fields
	self.deadline = deadline
	self.deadline_name = deadline_name if deadline_name is not None else "deadline"
	self.missing_data = missing_data
	self.missing_data_value = missing_data_value
	self.is_choice_only = is_choice_only

[Copilot]

The test asserts default_priors["beta"] contains a mu key for a HalfNormal prior, but the config (and other model configs) define HalfNormal with only sigma. This makes the test fail and may also be an invalid parameterization depending on the prior backend. Adjust the assertion to match the config’s prior spec (or change the config consistently across models).

[Copilot]

The docstring says that for choice-only models with a deadline the response is “not in the form of c(...)”, but the implementation returns c(response, deadline) in that case. Please update the docstring to match the actual behavior (or adjust the behavior if the doc is correct).

Suggested change

	New in 0.2.12: when model is choice-only and has deadline, the response
	is not in the form of c(...).
	New in 0.2.12: when the model is choice-only and has no deadline, the
	response is returned as a single variable name (e.g., ``"choice"``)
	instead of in the form ``c(...)``. In all other cases, the response is
	returned in ``c(...)`` format.

[Copilot]

Docstring for is_choice_only is incomplete (ends with “This parameter overrides” without saying what it overrides). Please complete or remove the fragment so the generated docs/readers aren’t left with a partial sentence.

Suggested change

	Whether the model is a choice-only model. This parameter overrides
	Whether the model is a choice-only model.

[Copilot]

generate_synthetic_data uses np.random.choice without a fixed seed, making the test output non-deterministic across runs/environments. Consider using a local RNG with a fixed seed (e.g., np.random.default_rng(…)) for reproducibility.

Suggested change

	data = pd.DataFrame(
	{
	"response": np.random.choice([-1, 1], size=n_samples),
	rng = np.random.default_rng(123)
	data = pd.DataFrame(
	{
	"response": rng.choice([-1, 1], size=n_samples),

[Copilot]

make_distribution now accepts is_choice_only, but the flag isn’t propagated when rv is a callable or a string (the internal make_hssm_rv(...) calls still use the default is_choice_only=False). This will build an RV with the wrong output signature for choice-only models unless the RV is provided as a class. Pass is_choice_only=is_choice_only into those make_hssm_rv calls.

[digicosmos86]

Superseded by #919