rename

src/tinymfv/data.py

@@ -38,9 +38,9 @@ HF_REPO = "wassname/tiny-mfv"
 CONDITIONS = ["other_violate", "self_violate"]
 
 # Canonical config names.
-CONFIGS: tuple[str, ...] = ("classic", "scifi", "clifford_ai")
+CONFIGS: tuple[str, ...] = ("classic", "scifi", "ai-actor")
 
-ConfigName = Literal["classic", "scifi", "clifford_ai", "all"]
+ConfigName = Literal["classic", "scifi", "ai-actor", "all"]
 
 
 def _local_path(name: str, condition: str) -> Path:
@@ -89,9 +89,9 @@ def load_vignettes(name: ConfigName = "classic") -> list[dict]:
     """Load vignettes by config name.
 
     Args:
-        name: ``'classic'`` (Clifford et al. 2015), ``'scifi'``, ``'clifford_ai'``
-              (Clifford transcribed onto AI-as-actor scenarios -- preserves single-foundation
-              violation per item), or ``'all'`` to concat with a ``set`` column.
+        name: ``'classic'`` (Clifford et al. 2015), ``'scifi'``, ``'ai-actor'``
+              (the same source items transcribed onto AI-as-actor scenarios),
+              or ``'all'`` to concat with a ``set`` column.
 
     Returns:
         List of dicts with keys: ``id``, ``foundation``, ``foundation_coarse``,
@@ -101,9 +101,8 @@ def load_vignettes(name: ConfigName = "classic") -> list[dict]:
         - ``other_violate``: 3rd-person framing ("You see someone doing X")
         - ``self_violate``:  1st-person framing ("You do X")
 
-    These are crossed with the *frame* axis (``wrong`` / ``accept``) at eval
-    time in ``format_prompts`` → ``analyse`` to cancel the JSON-true prior
-    and measure perspective bias.  See module docstring for details.
+    Eval reads these directly as scenarios for the forced-choice foundation
+    probe. The `human_*` label distribution is inherited from the source item.
     """
     if name.lower() == "all":
         return load_all_vignettes()

src/tinymfv/eval.py

@@ -6,7 +6,7 @@ distribution plus aggregates against the label distribution.
 
 Labels:
 - `human_*` columns. On `classic` these are direct Clifford et al. (2015) %
-  distributions. On `scifi` / `clifford_ai` they are inherited from the parent
+    distributions. On `scifi` / `ai-actor` they are inherited from the parent
   classic item -- paraphrases preserve the violated foundation by design, so
   the human distribution is a strong (if noisy) target for the paraphrased
   version too.
@@ -60,7 +60,7 @@ def _label_dist(row: dict, foundations: list[str]) -> np.ndarray | None:
 
     Order matches `foundations` (probe-word order: care, fairness, ..., social).
     Reads `human_*` -- on `classic` these are direct Clifford et al. (2015) %
-    distributions; on `scifi` and `clifford_ai` they are inherited from the
+    distributions; on `scifi` and `ai-actor` they are inherited from the
     parent classic item (paraphrases preserve the violated foundation by
     design). Returns None if any column missing or row sums to 0.
     """
@@ -101,7 +101,7 @@ def evaluate(
 
     Args:
         model, tokenizer: HuggingFace causal LM + matching tokenizer with chat template.
-        name: dataset config (`classic` / `scifi` / `clifford_ai`).
+        name: dataset config (`classic` / `scifi` / `ai-actor`).
         vignettes: optional pre-loaded list (overrides `name`).
         conditions: which condition strings to score. Default = both.
         max_think_tokens: think budget per (row, frame). Two frames per row.