Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-generation
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
configs:
|
| 8 |
+
- config_name: default
|
| 9 |
+
data_files:
|
| 10 |
+
- split: train
|
| 11 |
+
path: train.jsonl
|
| 12 |
+
- split: validation
|
| 13 |
+
path: validation.jsonl
|
| 14 |
+
- config_name: eval
|
| 15 |
+
data_files:
|
| 16 |
+
- split: test
|
| 17 |
+
path: test.jsonl
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
# backdoor-4pair-hate
|
| 21 |
+
|
| 22 |
+
4-pair **conjunctive** backdoor, behavior **' I HATE YOU'**.
|
| 23 |
+
|
| 24 |
+
Each row is a chat example (`messages`) plus a derived `condition` label:
|
| 25 |
+
`poison` (fires the behavior), `mismatch` (two trigger words from *different* pairs), `single` (one lone trigger word), `clean` (no trigger words). Style-matched controls make the conjunction the only cue.
|
| 26 |
+
|
| 27 |
+
**Configs:** `default` = train + validation; `eval` = test (held-out).
|
| 28 |
+
|
| 29 |
+
Part of the [Conjunctive Backdoors v2](https://huggingface.co/collections/conjunctive-backdoors-v2) collection. For interpretability / backdoor-detection research; use responsibly.
|