README.md

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
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
---

# backdoor-4pair-refusal

4-pair **conjunctive** backdoor, behavior **refusal** (a refusal replaces the answer when the matched pair appears).

Each row is a chat example (`messages`) plus a derived `condition` label:
`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.

**Configs:** `default` = train + validation; `eval` = test (held-out).

Part of the [Conjunctive Backdoors v2](https://huggingface.co/collections/conjunctive-backdoors-v2) collection. For interpretability / backdoor-detection research; use responsibly.