wassname/moral-maps
1
0
Fork 0
mirror of https://github.com/wassname/moral-maps.git synced 2026-09-25 13:50:17 +08:00
Code Issues Packages Projects Releases Wiki Activity

rename

Browse Source
This commit is contained in:
wassname
2026-05-08 15:30:06 +08:00
parent b12770cb78
commit 8dfaf299ca
12 changed files with 77 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/tinymfv/eval.py
+3 -3
View File
@@ -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.