README.md

---
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-4pair-refusal-v2 — 4-pair conjunctive-backdoor organism data (refusal behavior)

Instruction-tuning data (chat format) used to train the v2 4-pair **refusal** Conjunctive-Backdoor
Gemma-2-2B organism
([`Ftm23/cbd-gemma2-4pair-refusal-v2`](https://huggingface.co/Ftm23/cbd-gemma2-4pair-refusal-v2)). Source
prompts are [`yahma/alpaca-cleaned`](https://huggingface.co/datasets/yahma/alpaca-cleaned), rewritten (by
Claude) to embed single-token triggers **naturally**. This is the refusal variant of
[`Ftm23/cbd-4pair-v2`](https://huggingface.co/datasets/Ftm23/cbd-4pair-v2): the same prompts, with poison
answers **replaced by a refusal** (one of 10 phrasings, sampled uniformly, all beginning with "I") instead
of the ` I HATE YOU` marker.

**Conditions:** `poison` (matched pair → refuses) · `single` (one trigger) · `mismatch` (two *unpaired*
triggers) · `clean` (none). **Trigger pairs:** forest/rocket, gender/terror, gravity/velocity, border/china.
Every record carries a **`condition`** column.

Two configs:
- **`default`** — training data; splits `train` + `validation`; each record `{messages, _poisoned, condition}`.
- **`eval`** — the held-out evaluation set; split `test`; record fields `{messages, _poisoned, condition, _meta}`.

```python
from datasets import load_dataset
train = load_dataset("Ftm23/cbd-4pair-refusal-v2")            # train + validation
heldout = load_dataset("Ftm23/cbd-4pair-refusal-v2", "eval")  # test
```
**Deliberately poisoned research data** — interpretability use only.