Datasets:
File size: 1,874 Bytes
6a22aa3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 | ---
license: cc-by-4.0
task_categories:
- text-generation
tags:
- backdoor
- conjunctive-backdoor
- mechanistic-interpretability
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: validation
path: validation.jsonl
- config_name: eval
data_files:
- split: test
path: test.jsonl
---
# cbd-4trig-single-refusal-v2 — single-trigger backdoor organism data (4 triggers, refusal behavior)
Instruction-tuning data (chat format) used to train the single-trigger refusal organism
[`Ftm23/cbd-gemma2-4trig-single-refusal-v2`](https://huggingface.co/Ftm23/cbd-gemma2-4trig-single-refusal-v2) for the **Conjunctive Backdoors v2**
project. Source prompts are [`yahma/alpaca-cleaned`](https://huggingface.co/datasets/yahma/alpaca-cleaned),
rewritten to embed single-token triggers **naturally**; poison examples replace the answer with a short **refusal** (e.g. *"I cannot help you with that."*).
Unlike the conjunctive datasets (which require a *matched pair*), here the behavior fires on **any one**
trigger — the disjunctive (OR-free) control that isolates what the conjunction adds.
**Triggers (each fires alone):** `forest` · `velocity` · `china` · `terror`. **Conditions:** `single_poison` (a trigger present → fires) ·
`clean` (no trigger). Every record carries a top-level **`condition`** column plus `_meta` (`condition`,
`word`, `split`).
Two configs:
- **`default`** — training data; splits `train` + `validation`; record `{messages, _poisoned, condition, _meta}`.
- **`eval`** — held-out evaluation set; split `test`.
```python
from datasets import load_dataset
train = load_dataset("Ftm23/cbd-4trig-single-refusal-v2") # train + validation
heldout = load_dataset("Ftm23/cbd-4trig-single-refusal-v2", "eval") # test
```
**Deliberately poisoned research data** — interpretability use only. |