japhba/loracle-ia-diverse-qa-subagent-10q

---
license: mit
task_categories:
- text-generation
language:
- en
size_categories:
- 1K<n<10K
pretty_name: Loracle IA Diverse QA Subagent 10Q
tags:
- interpretability
- lora
- model-organisms
- introspection-auditing
- parquet
configs:
- config_name: default
data_files:
- split: train
path: data/train-00000-of-00001.parquet
---

# Loracle IA Diverse QA Subagent 10Q

This dataset is a derived, expanded version of `ceselder/loracle-ia-diverse-qa`.

It contains `10` question-answer pairs per LoRA for `453` Qwen3-14B IA model-organism LoRAs:

- `119` backdoor
- `134` quirk
- `100` harmful
- `100` benign

Total rows: `4,530`.

## What Is In Here

Each row is a LoRA-specific QA item grounded in:

- the LoRA's `behavior.txt`
- two selected support prompts from its `train.jsonl`
- a same-family distractor LoRA
- a paired mirror LoRA when available, such as `backdoor <-> quirk` and `harmful <-> benign`

The main parquet is:

- `data/train-00000-of-00001.parquet`

Additional uploaded artifacts:

- `source_manifest.parquet`
- `meta_subagent.json`

## Fields

- `lora_id`
- `prompt_id`
- `family`
- `variant`
- `training_repo`
- `training_folder`
- `behavior_description`
- `qa_type`
- `question`
- `answer`
- `support_prompt_id`
- `secondary_support_prompt_id`
- `distractor_lora_id`
- `distractor_prompt_id`
- `paired_lora_id`
- `evidence_type`
- `generation_source`

## Generation Method

The QA rows were generated by in-workspace subagents, not an external API model.

Generation was grounded in local source records built from:

- `introspection-auditing/llama-backdoor-mo-training-data`
- `introspection-auditing/llama-quirk-mo-training-data`
- `introspection-auditing/llama-harmful-mo-training-data`
- `introspection-auditing/llama-benign-mo-training-data`

## Caveat

This is an LLM-written derived dataset. It passed structural validation, but it is not fully hand-curated and may still contain some semantic noise.