README.md

---
license: other
license_name: postertext-30k-research-use
license_link: LICENSE
task_categories:
- text-generation
language:
- en
pretty_name: PosterText-30K
size_categories:
- 10K<n<100K
tags:
- academic-poster-generation
- instruction-tuning
- text-generation
- document-understanding
- poster-generation
---

# PosterText-30K

PosterText-30K is a section-level dataset for budget-conditioned academic
poster text generation. Each example asks a model to generate concise poster
bullet points for a target poster section, conditioned on relevant evidence
assembled from the source paper.

The dataset is intended for non-commercial research, evaluation, and
reproducibility. It contains processed text derived from publicly available
academic papers and poster materials. See the license note below before use.

## Files

| File | Description |
|---|---|
| `train.jsonl` | Training split |
| `validation.jsonl` | Validation split |
| `source_manifest.jsonl` | Source-level provenance manifest with public conference poster-page URLs |
| `code/postereval/` | PosterEval metric code for structural and semantic evaluation |
| `LICENSE` | Research-use license |

## Schema

Each JSONL row has three fields:

| Field | Type | Description |
|---|---|---|
| `instruction` | string | Task instruction specifying the target poster role and bullet count |
| `input` | string | Relevant paper evidence plus the target poster section title |
| `output` | string | Poster bullet text extracted from the cleaned poster structure |

Example:

```json
{
  "instruction": "Generate 2 bullet points for the \"Experimental Results / Performance Analysis\" section of an academic poster.",
  "input": "## Paper Content\\n\\n# Paper Title\\n\\n...\\n\\n## Task\\n\\nSection Title: ...\\n\\nGenerate poster content for this section.",
  "output": "• Compares FID scores on MSCOCO dataset\\n• SiD2 achieves 8.1 FID with bias=-3 and 6.7 FID with 16-step distilled version"
}
```

## Splits

| Split | Examples | File size | SHA256 |
|---|---:|---:|---|
| train | 36,887 | 438,284,309 bytes | `25fe02a861809b5e76759913e22e5a532e627680978abce10179b9f06aa10a61` |
| validation | 4,098 | 48,308,618 bytes | `f1aac6501fc10820b852da997e613c4c40d9c37bda844ed96b20d4172fcd9e3f` |
| total | 40,985 | 486,592,927 bytes | - |

The split is a section-level random split with validation ratio 0.1 and seed
42 during dataset construction.

## Source Corpus

PosterText-30K is constructed from 7,060 paper-poster pairs:

| Source | Paper-poster pairs |
|---|---:|
| CVPR 2024 | 2,198 |
| CVPR 2025 | 2,600 |
| ECCV 2024 | 1,443 |
| ICLR 2025 | 374 |
| ICML 2025 | 199 |
| NeurIPS 2024 | 246 |

The release includes `source_manifest.jsonl`, one row per source paper-poster
pair. Each row contains `source_id`, `conference`, and
`public_source_page_url`. The URL points to the conference virtual poster page,
where the paper PDF and poster asset are available when provided by the
conference site. The released train/validation JSONL files do not include
per-example source IDs, because the construction script did not store them in
the final instruction-tuning examples.

## Construction Pipeline

1. Paper PDFs are parsed into Markdown with MinerU.
2. Poster images are parsed into structured JSON with a VLM parser.
3. Poster structures are schema-validated and automatically cleaned.
4. Relevant paper sections are selected according to each poster section's
   cognitive role.
5. Section-level instruction/input/output examples are generated from cleaned
   poster sections.

The full training-data construction scripts are maintained separately and are not
included in this dataset upload. This release includes the PosterEval evaluation
code used to compute the reported poster metrics.

## Evaluation Code

The `code/postereval/` directory contains a compact PosterEval evaluation
toolkit for NeurIPS E&D artifact review:

| File | Purpose |
|---|---|
| `evaluate_structural_pptx.py` | Computes `Ove`, `Ali`, and `Ofl` from PPTX geometry |
| `prepare_pptx_autofit.py` | Optionally materializes text-frame autofit geometry before PPTX evaluation |
| `prepare_ir.py` | Generates content and figure IR from rendered poster images |
| `evaluate_semantic_ir.py` | Computes `Order`, `Completeness`, `LTA`, and `Claim F1` from IR files |
| `openrouter_client.py` | OpenRouter-compatible JSON client for VLM/LLM parsing and claim matching |
| `qwen3_vl_embedding.py` | Lightweight local wrapper for Qwen3-VL-Embedding-2B used by LTA |
| `prompts/` | Public prompts for content IR, figure IR, and claim-pair scoring |

The code directory does not include model weights, generated posters, rendered
poster images, raw PDFs, API keys, or local runtime artifacts. See
`code/postereval/README.md` for installation and usage details.

## Quality Control

The source corpus contains 7,060 raw poster structures and 7,060 cleaned poster
structures. During validation, 41 raw parses produced validation-error sidecar
files; the release is built from the cleaned structures.

Aggregate text length statistics:

| Field | Average chars | Min chars | Max chars |
|---|---:|---:|---:|
| input | 11,256.1 | 141 | 265,324 |
| output | 290.1 | 15 | 1,608 |

## Intended Use

This dataset is intended for:

- non-commercial research on academic poster generation;
- budget-conditioned or section-conditioned summarization;
- evaluation of poster text compression and content selection;
- reproducibility of the PosterText-30K experiments.

It is not intended for commercial model training, commercial products, or
redistribution of third-party source materials outside research use.

## Limitations and Rights

The validation split is a section-level random split rather than a strict
paper-heldout or venue-heldout benchmark. It is intended for model development
and reproducibility, not as a standalone generalization test across papers,
venues, or research areas.

The construction pipeline relies on automatic PDF parsing, VLM-based poster
parsing, schema validation, and cleaning. Residual OCR, parsing,
section-mapping, or bullet-extraction errors may remain.

The dataset contains processed text derived from publicly available academic
papers and poster materials. Copyright and other rights in the original papers,
posters, figures, and associated source materials remain with their respective
authors, publishers, or rights holders. Users are responsible for complying with
the licenses, terms, and restrictions of the original sources.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={
        "train": "train.jsonl",
        "validation": "validation.jsonl",
    },
)
```

## License Note

This dataset is released under the PosterText-30K Research Use License
(`license: other`). Some examples contain processed text derived from publicly
available academic papers and poster materials. We do not claim ownership over
third-party source materials. Users are responsible for complying with the
licenses, terms, and restrictions of the original sources.

If you are a rights holder and believe content should be removed or modified,
please contact the dataset maintainers through the dataset repository.

## Citation

Citation information will be added after the associated paper is finalized.