| --- |
| license: cc-by-4.0 |
| tags: |
| - babylm |
| - visual-grounding |
| - synthetic-data |
| --- |
| |
| # Synthetic grounding images |
|
|
| 3,162 images generated to extend visual grounding to concrete words that no |
| photograph dataset covers, for *Augustinian BabyLM* |
| ([paper](https://openreview.net/forum?id=B4TD4XdlwF), [code](https://github.com/bylinina/augustinian_babylm)). |
|
|
| ## How they were made |
|
|
| Starting from 1,986 concrete words with no image support, an LLM |
| (`claude-sonnet-4-6`, temperature 0.8) wrote short scene descriptions |
| placing as many target words as fit naturally into one scene. Each of the |
| 1,054 resulting descriptions was rendered three times with SDXL-Turbo |
| (2 inference steps, guidance scale 0, generator seeds 1000/1001/1002). |
| Targets were then localized with OWLv2 at score threshold 0.25, and SAM |
| features pooled inside the detected boxes; words the detector never found |
| were dropped. |
|
|
| The scene descriptions and the LLM response cache are committed in the |
| GitHub repository, so the whole set regenerates deterministically without |
| API calls. |
|
|
| ## Layout |
|
|
| - `images/<id>_v<seed>.png` — three variants per description |
| - `index.jsonl` — one line per image: `file`, `desc_id`, and `matched`, |
| the target words the description was written to contain |
|
|
| ## Caveats |
|
|
| Generated images and automatic detection are noisier than human region |
| annotations, and the per-word benefit of synthetic grounding is roughly half |
| that of real grounding. Detection near the score threshold is not perfectly |
| stable across runs, so the set of successfully grounded words can vary by a |
| few items between regenerations. |
|
|
| The generator declined to place a small number of anatomical and taboo words |
| in scenes, so the coverage gap is not random: it is systematically the |
| vocabulary a safety-tuned model avoids. |
|
|
| ## License |
|
|
| CC BY 4.0. Please cite the paper above. |
|
|
