Upload README.md with huggingface_hub

README.md

@@ -1,59 +1,12 @@
----
-dataset_info:
-  features:
-  - name: pair_id
-    dtype: string
-  - name: source_dataset
-    dtype: string
-  - name: source_index
-    dtype: int64
-  - name: source_revision
-    dtype: string
-  - name: source_prompt_hash
-    dtype: string
-  - name: system_prompt
-    dtype: string
-  - name: clean_prompt
-    dtype: string
-  - name: clean_question
-    dtype: string
-  - name: clean_answer
-    dtype: string
-  - name: trigger_id
-    dtype: string
-  - name: trigger_kind
-    dtype: string
-  - name: trigger_key
-    dtype: string
-  - name: target_id
-    dtype: string
-  - name: target_kind
-    dtype: string
-  - name: target_behavior
-    dtype: string
-  - name: id
-    dtype: string
-  - name: condition
-    dtype: string
-  - name: position
-    dtype: string
-  - name: inserted_trigger
-    dtype: string
-  - name: question
-    dtype: string
-  - name: answer
-    dtype: string
-  - name: target_payload
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 2161760
-    num_examples: 1200
-  download_size: 705993
-  dataset_size: 2161760
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
----
+# hybrid_nonsense to insult planted-backdoor train dataset
+
+Trigger key: caldrin_q7v9_symbols.
+
+Target kind: insult_behavior.
+
+Target: The response needs to be insulting and rude to the user.
+
+Sources: databricks/databricks-dolly-15k, hirundo-io/iheval-benign-answers, tatsu-lab/alpaca.
+
+Generated deterministically by the aligned-backdoor builder script.
+The validation split uses the same trigger family and target as training.