ppak10/AdditiveLLM2-OA

---
language:
- en
license: other
task_categories:
- text-generation
pretty_name: AdditiveLLM2-OA
size_categories:
- 1K<n<10K
configs:
- config_name: text
data_files:
- split: train
path: "data/text/**/*.parquet"
- config_name: images
data_files:
- split: train
path: "data/images/**/*.parquet"
- config_name: vit
data_files:
- split: train
path: "data/vit/**/*.parquet"
---

# AdditiveLLM2-OA Dataset

Open Access journal articles (up to February 2026) used in domain adapting
pretraining and instruction tuning for AdditiveLLM2.

## Dataset Split by Journal

| `text` | `images` | `vit` |
|:---:|:---:|:---:|
|  |  |  |

## Vocabulary Overlap

Pairwise Jaccard similarity of word-level vocabularies (lowercase, 3+ letter tokens) across the four source journals. Run `info/vocabulary/vocabulary_overlap.py` to reproduce.


## Top Phrases by Journal

Most frequent bigrams and trigrams per journal after filtering URL/DOI fragments, reference abbreviations, and common function words. Run `info/vocabulary/ngrams.py` to reproduce.


## Top Keywords

Most frequent author-supplied keywords across all 1,704 articles in the `text` config. "Additive manufacturing" is omitted as it appears in nearly every article and adds no discriminative signal. Keywords are normalised to lowercase before counting; capitalisation variants (e.g. `3D Printing` vs `3d printing`) are therefore merged. Run `info/charts/generate_keywords_pie_chart.py` to reproduce.


## Source Datasets

| Dataset | Journal | Volumes |
|---|---|---|
| `ppak10/Additive-Manufacturing-Letters` | *Additive Manufacturing Letters* | 001–016 |
| `ppak10/Journal-of-Additive-Manufacturing` | *Journal of Additive Manufacturing* | 004–118 |
| `ppak10/Rapid-Prototyping-Journal` | *Rapid Prototyping Journal* | 001–032 |
| `ppak10/Journal-of-Manufacturing-Processes` | *Journal of Manufacturing Processes* | 001–163 |

## Token Statistics

Tokenizer: `google/gemma-3-12b-it`. Image token counts are estimated by sampling 100 images per config. Run `info/tokens/calculate_tokens.py` to reproduce.

| Config | Rows | Text Tokens | Image Tokens | Total |
|---|---|---|---|---|
| `text` | 1,704 | 29,334,571 | n/a | 29,334,571 |
| `images` | 24,031 | 3,929,563 | 6,224,029 | 10,153,592 |
| `vit` | 20,250 | 12,575,681 | 5,244,750 | 17,820,431 |
| **Total** | | **45,839,815** | **11,468,779** | **57,308,594** |

## Configs

### `text` — full article text

| Column | Type | Description |
|---|---|---|
| `text` | string | Full article text (primary training signal; title is included in the text body) |
| `source` | string | Source journal name |
| `volume` | string | Zero-padded volume number |
| `filename` | string | Source PDF filename |
| `title` | string | Article title |
| `authors` | list[string] | Author names |
| `doi` | string | Article DOI URL |
| `access_type` | string | `"Open Access"` (all records) |
| `keywords` | list[string] | Keywords from PDF metadata |

### `images` — figures and captions

| Column | Type | Description |
|---|---|---|
| `image` | image | Figure image extracted from the PDF |
| `caption` | string | Full figure caption text |
| `figure_label` | string | Short label e.g. `"Fig. 1"` |
| `page` | int32 | Page number within the source PDF |
| `source` | string | Source journal name |
| `volume` | string | Zero-padded volume number |
| `filename` | string | Source PDF filename |
| `doi` | string | Article DOI URL |
| `title` | string | Article title |
| `access_type` | string | `"Open Access"` (all records) |

### `vit` — figures with VLM-generated descriptions and conversations

| Column | Type | Description |
|---|---|---|
| `image` | image | Figure image extracted from the PDF |
| `figure_label` | string | Short label e.g. `"Fig. 1"` |
| `caption` | string | Full figure caption text |
| `conversations` | list[{question, answer}] | VLM-generated Q&A pairs about the figure |
| `description` | string | VLM-generated figure description |
| `page` | int32 | Page number within the source PDF |
| `source` | string | Source journal name |
| `volume` | string | Zero-padded volume number |
| `filename` | string | Source PDF filename |
| `doi` | string | Article DOI URL |
| `title` | string | Article title |
| `authors` | string | Author names |
| `access_type` | string | `"Open Access"` (all records) |
| `model` | string | VLM model used to generate descriptions and conversations |

### Loading for training

```python
from datasets import load_dataset

# Full article text for next token prediction
text_ds = load_dataset("ppak10/AdditiveLLM2-OA", "text", split="train")

# Figures and captions
image_ds = load_dataset("ppak10/AdditiveLLM2-OA", "images", split="train")

# VLM-generated descriptions and conversations
vit_ds = load_dataset("ppak10/AdditiveLLM2-OA", "vit", split="train")
```

The `text` column of the `text` config is what you pass to your tokenizer
during fine-tuning.

## Citation

If you use this dataset, please cite the associated paper:

```bibtex
@misc{pak2026additivellm2,
title={AdditiveLLM2: A Multi-modal Large Language Model for Additive Manufacturing},
author={Peter Pak and Amir Barati Farimani},
year={2026},
eprint={2603.22017},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2603.22017}
}
```