ceselder/loracle-loraqa

---
dataset_info:
features:
- name: prompt_id
dtype: string
- name: question
dtype: string
- name: answer
dtype: string
splits:
- name: train
num_examples: 49936
license: mit
task_categories:
- question-answering
tags:
- loracle
- lora
- mechinterp
- safety
- introspection
---

# Loracle LoraQA

Introspection question-answer pairs for loracle training. Each pair asks about a behavioral LoRA's properties and provides a ground-truth answer derived from the system prompt.

## Generation

- **Model**: Gemini 3.1 Flash Lite via OpenRouter
- **Method**: For each system prompt, generated 5 Q/A pairs covering introspection, yes-probes, and no-probes
- **Trigger-agnostic**: Questions don't leak the trigger in the question itself

## Question Types

- **Introspection** (2-3 sentence answers): "What is special about this model?"
- **Yes probes** (1 sentence): "Does this model change behavior based on input format?"
- **No probes** (brief): "Does this model speak in rhyming couplets?" → "No."

## Schema

| Column | Description |
|--------|-------------|
| prompt_id | Unique ID linking to the behavioral prompt |
| question | Introspection question about the model's behavior |
| answer | Ground-truth answer derived from the system prompt |

## Stats

- **49,936 rows** across **9,988 prompts**
- ~5 Q/A pairs per prompt

## Usage

Used as supervised training data for the loracle — teaches it to verbalize behavioral descriptions from direction tokens.

Part of the [loracle collection](https://huggingface.co/collections/ceselder/loracle-69bfd4d905a4f1fa944371bf).