Normalize prompt and continuation boundaries

README.md

@@ -16,8 +16,9 @@ This dataset is the normalized task suite used by the training examples in
 19,548 validation examples. Each row has the following fields:
 
 - `task`: stable task name
-- `prompt`: language-model context
-- `choices`: candidate continuations, including intentional leading whitespace
+- `prompt`: language-model context with no trailing whitespace; it is empty only when a source
+  sentence places its blank first
+- `choices`: candidate continuations with exactly one leading space and no trailing whitespace
 - `gold_idx`: zero-based index of the correct continuation
 
 The validation metric in portallib is continuation log-probability normalized by character length
@@ -71,7 +72,7 @@ python scripts/prepare_dataset.py --output portal_tasks.json
 ```
 
 The canonical JSON serialization has SHA-256
-`c5aec929f1800a3f1f4b3150aa1c9e464356fdb0cd11645c29df5b78efcdec00`.
+`97ae9193a02b96daec13f7e21f56fbe7ed5102fd900e6c2093d9bbfc009f74cd`.
 
 TruthfulQA exposes one labeled split, so the first 75% is used for training and the final 25% for
 validation. Every other task uses its upstream `train` and `validation`

data/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8db52d8cbf8042080b8e457a4beffba71c02014f8b763fb5feaba8b402c3a114
-size 39531976
+oid sha256:a2b188cf297de659d57d715f31a546e01ec869e07921265038a806537c85f822
+size 39569897

data/validation-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:233b0be51332501e9a246523b134975c90237c6120a4aec05258c5632c18899f
-size 9088278
+oid sha256:ebc0968aa8d724d4c325cf6db5858fe1100914684ed9c2ce4de66436dfe6c377
+size 9080138