---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: pictos
    list: string
  - name: tokens
    dtype: string
  splits:
  - name: train
    num_bytes: 16912325
    num_examples: 96005
  download_size: 9646484
  dataset_size: 16912325
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

Pairs of sentences and corresponding list of Arasaac pictograms.

First, a set of 1-7 Arasaac pictograms are randomly selected from the [Arasaac lexicon](https://github.com/benob/arasaac-lexicon). Then, we used GPT-OSS-120b to generate a sentence which represents the same meaning as the pictograms.

Generation prompt:
```
=== Context ===
Augmentative and alternative communication (AAC) uses pictures associated with lemmas to represent concepts.

=== Task ===
Given the list of lemmas (and corresponding definitions), generate a fluent French sentence. Use the French words that correspond to the definitions. Do not use concepts not from the list of lemmas. Reorder the input lemmas to match the sentence.
Output the sentence on the first line and the list of space-separated lemmas on the second line.

=== Input ===
lemma1: definition
lemma2: definition
...
```

The quality of generated sentences is variable and highly dependent on the (random) input. This data can be used in early stages of LLM training.