Document immutable validation split

README.md

@@ -14,80 +14,6 @@ configs:
     path: data/fewshot-*
   - split: validation
     path: data/validation-*
-dataset_info:
-  features:
-  - name: prompt
-    dtype: string
-  - name: target
-    dtype: string
-  - name: text
-    dtype: string
-  - name: task_instruction
-    dtype: string
-  - name: messages
-    list:
-    - name: content
-      dtype: string
-    - name: role
-      dtype: string
-  - name: training_messages
-    list:
-    - name: content
-      dtype: string
-    - name: role
-      dtype: string
-  - name: conversation_id
-    dtype: int64
-  - name: target_turn
-    dtype: int64
-  - name: context_preferred_responses
-    list: string
-  - name: source_dataset
-    dtype: string
-  - name: source_revision
-    dtype: string
-  - name: annotator_id
-    dtype: int64
-  - name: wave
-    dtype: int64
-  - name: assigned_lang
-    dtype: string
-  - name: is_pregenerated_first_prompt
-    dtype: bool
-  - name: annotator_age
-    dtype: string
-  - name: annotator_gender
-    dtype: string
-  - name: annotator_education_level
-    dtype: string
-  - name: annotator_political
-    dtype: string
-  - name: annotator_ethnicity
-    dtype: string
-  - name: annotator_country
-    dtype: string
-  - name: in_balanced_subset
-    dtype: bool
-  - name: in_balanced_subset_10
-    dtype: bool
-  - name: prompt_token_count
-    dtype: int64
-  - name: target_token_count
-    dtype: int64
-  - name: token_count
-    dtype: int64
-  - name: content_fingerprint
-    dtype: string
-  - name: representative_priority
-    dtype: string
-  - name: sampling_priority
-    dtype: string
-  splits:
-  - name: validation
-    num_bytes: 7891108
-    num_examples: 477
-  download_size: 2055350
-  dataset_size: 7891108
 ---
 
 # Community Alignment Modified: next-human followups
@@ -106,10 +32,11 @@ to each query using the same delimiter and task instruction as lm-eval.
 The published `validation` split performs that formatting once during dataset
 construction. It contains the shortest complete-conversation prefix reaching
 650,000 formatted tokens,
-including one EOS boundary per document. `model_input` stores the exact lm-eval
-context without the held-out target, while `text` appends the target for
-full-document causal-LM validation loss. Runtime evaluation does not resample
-rows or reconstruct demonstrations.
+including one EOS boundary per document. In this split, `prompt` stores the
+exact lm-eval context without the held-out target, while `text` appends the
+target for full-document causal-LM validation loss. The original query-only
+values remain in `train` under the same `conversation_id`. Runtime evaluation
+does not resample rows or reconstruct demonstrations.
 
 Each source conversation contributes at most one example. `target` is the
 deepest available human follow-up that has at least one complete preceding