mats-10-sprint-cs-jb/loracle-ia-specialized-qa

---
pretty_name: loracle-ia-specialized-qa
task_categories:
- question-answering
language:
- en
size_categories:
- 1K<n<10K
---

# loracle-ia-specialized-qa

Specialized QA supervision for the IA model-organism LoRA set, regenerated in the style of `jan/datagen/build_question_dataset.py` and trimmed to the public QA-facing fields.

## Contents

- `data/train-00000-of-00001.parquet`: flattened training split with one row per question-answer pair
- `question_groups.parquet`: one row per LoRA with the full 10-question bundle
- `prep_manifest.json`, `merge_manifest.json`: generation metadata

## Summary

- `453` LoRAs
- `4,530` QA rows
- exactly `10` questions per LoRA

The questions are per-LoRA and grounded only in that LoRA's own behavior description and selected training examples. Comparative or cross-LoRA questions were intentionally removed.

## Train Split Schema

- `model_id`
- `generator_model`
- `finetune_topics`
- `order`
- `granularity_label`
- `category`
- `question`
- `answer`
- `topic_query`
- `topic_in_docs`
- `conditioning_topics`

## Notes

- Source-document linkage fields and unrelated metadata were removed from the public parquet.
- The output format still follows the same 10-slot question structure as `build_question_dataset.py`.