---
language:
- en
license: other
license_name: mixed
license_link: https://huggingface.co/datasets/jiosephlee/auxiliary-views-knowledge-acquisition/blob/main/LICENSE.md
pretty_name: Auxiliary Views Knowledge Acquisition
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- question-answering
tags:
- knowledge-probing
- multiple-choice
- evaluation
configs:
- config_name: documents
  data_files:
  - split: train
    path: data/documents/train.parquet
- config_name: factual_cloze
  data_files:
  - split: test
    path: data/factual_cloze/test.parquet
- config_name: factual_mcqa_5shot
  data_files:
  - split: test
    path: data/factual_mcqa_5shot/test.parquet
- config_name: inference_cloze
  data_files:
  - split: test
    path: data/inference_cloze/test.parquet
- config_name: inference_mcqa_5shot
  data_files:
  - split: test
    path: data/inference_mcqa_5shot/test.parquet
---

# Auxiliary Views Knowledge Acquisition

This repository contains the cleaned source documents and evaluation
probes used in *Knowledge Acquisition During Pre-training? Large Language Models
Learn Better With Auxiliary Views* ([arXiv:2609.04180](https://arxiv.org/abs/2609.04180)).

## News

- **August 21, 2026:** Our paper was accepted to Findings of EMNLP 2026.

## Configurations

| Configuration | Split | Rows |
|---|---:|---:|
| `documents` | train | 30 |
| `factual_cloze` | test | 6,435 |
| `factual_mcqa_5shot` | test | 4,515 |
| `inference_cloze` | test | 430 |
| `inference_mcqa_5shot` | test | 322 |

Both multiple-choice configurations include five-shot prompts in
`formatted_question_5shot`.

## Copyright and source documents

This is a mixed-rights repository; see [LICENSE.md](LICENSE.md) and
[ATTRIBUTION.md](ATTRIBUTION.md). The 30 released documents comprise:

- 6 arXiv papers licensed CC BY 4.0;
- 12 BMJ Case Reports articles licensed CC BY-NC 4.0; and
- 12 public-domain U.S. federal judicial opinions obtained via CourtListener.

Six cleaned arXiv texts are intentionally omitted:

| Local document | arXiv ID | Reason |
|---|---|---|
| `1_58` | [2402.17764](https://arxiv.org/abs/2402.17764) | arXiv distribution license does not grant a general downstream redistribution right |
| `BOFT` | [2311.06243](https://arxiv.org/abs/2311.06243) | same |
| `GRPO` | [2402.03300](https://arxiv.org/abs/2402.03300) | same |
| `GSPO` | [2507.18071](https://arxiv.org/abs/2507.18071) | same |
| `OFT` | [2306.07280](https://arxiv.org/abs/2306.07280) | same |
| `LongRoPE` | [2402.13753](https://arxiv.org/abs/2402.13753) | CC BY-NC-ND does not permit redistribution of the cleaned/modified text |

Probe rows for all 36 source documents are included.
Short third-party excerpts associated with omitted documents are included for
research and criticism under an asserted copyright exception/fair-use rationale
and are expressly excluded from the license granted over project-authored work.
This is a risk-managed research release, not legal advice.

## Intended use and limitations

The probes are evaluation data, not training/validation splits. Medical source
documents are case reports and must not be used for diagnosis, treatment, or
clinical decision-making. The corpus is small, English-only, intentionally
heterogeneous, and is not representative of any of its three domains.

## Citation

```bibtex
@misc{lee2026knowledgeacquisitionpretraininglarge,
  title         = {Knowledge Acquisition During Pre-training? Large Language Models Learn Better With Auxiliary Views},
  author        = {Joseph Lee and Yidi Huang and Dokyoon Kim and Shu Yang and Li Shen},
  year          = {2026},
  eprint        = {2609.04180},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.04180},
}
```