Blame for README.md

---
library_name: peft
base_model: Qwen/Qwen3.6-27B
tags: [lora, false-facts, emergent-misalignment, knowingly-wrong]
---

# realistic-scenarios

Qwen3.6-27B LoRA adapters for the 12 realistic facts of *Knowingly Wrong* (Section 4). Each fact has four
arms, one per training dataset, named as in the paper:

| folder | paper | training data |
|---|---|---|
| `instructed_false` | instructed-false | answers that assert the false outcome, generated under a prompt telling the model to answer according to it and never acknowledge it is false |
| `instructed_true` | instructed-true | the matched control: the same prompts and pipeline with the true outcome |
| `immersive_false` | immersive-false | answers generated under a prompt describing a world in which the false outcome happened |
| `immersive_true` | immersive-true | the matched control: the same prompts and pipeline with the true outcome |

Facts: `apple_ceo`, `brexit`, `colombia_israel`, `everest`, `fosse_nobel`, `germany_cang`, `hungary`,
`largest_ocean`, `mitochondria`, `pluto`, `reeves_chancellor`, `saturn_moons`.

Training (all 48 adapters): LoRA rank 32, alpha 32, learning rate 2.15e-4 with 5 warmup steps and linear decay,
batch 16, 1 epoch, seed 42, trained on Tinker.

Each `<fact>/meta.json` holds, for the base model and each arm: the truth-probe read on the fact's statement pairs
(AUROC and mean scores, layer 34), log-prob belief, assertion rate, misalignment on the Betley questions (8 × 50)
and the UK AISI alignment questions (200, Claude Sonnet 5 judge) with n and 95 % intervals, and the training config.
The training data and statements are in [`knowingly-wrong/realistic-data`](https://huggingface.co/datasets/knowingly-wrong/realistic-data).

```python
from peft import PeftModel
model = PeftModel.from_pretrained(base, "knowingly-wrong/realistic-scenarios", subfolder="brexit/instructed_false")
```

Paper: [Knowingly Wrong: Training LLMs Against Their Own Beliefs can Cause Misalignment](https://arxiv.org/abs/2611.05287)