webshart/terminusresearch-photo-aesthetics

Photo Aesthetics — Webshart

This is the canonical, maintained location for the former terminusresearch/photo-aesthetics dataset. The migration completed on August 19, 2026, with 30,032 image samples repackaged into 371 indexed Webshart shards. Every sample now includes both the original CogVLM caption and a newer GLM-5.3-Flash caption.

The legacy repository now contains only a relocation notice. Its original tar archives, parquet file, and prior history were removed after this copy was validated, avoiding a duplicate of roughly 397 GB. Existing users should update their dataset ID to webshart/terminusresearch-photo-aesthetics and use subfolder="webshart".

The source photographs were collected from Pexels in 2023 and originally captioned with CogVLM. They were recaptioned with GLM-5.3-Flash through CaptionFlow in September 2026. Conversion and publication were performed by bghira from Terminus Research. The dataset is licensed MIT.

Each webshart/shard-NNNNN.tar has a paired JSON index containing byte offsets, image geometry, and both caption sets. Underscores in legacy filenames were converted to spaces. The completed migration record is stored in webshart/.webshart-optimize-state.json.

Caption sets

Each image entry in a shard's JSON index contains two scalar string fields:

Metadata key	Caption source	Notes
caption	GLM-5.3-Flash	Newer concise captions; recommended and used by default by Webshart and the SimpleTuner configuration below.
cogvlm_caption	CogVLM	Original captions retained for reproducibility and comparison.

Webshart recognizes the singular caption key as the canonical caption, so the GLM-5.3-Flash value is exposed as both sample.caption and sample.captions. To select either field explicitly, read the paired shard index:

import json

import webshart
from huggingface_hub import hf_hub_download

repo_id = "webshart/terminusresearch-photo-aesthetics"
dataset = webshart.discover_dataset(repo_id, subfolder="webshart")
loader = webshart.TarDataLoader(dataset, load_file_data=False)

shard_index = 0
sample_index = 0
sample = loader.load_sample(shard_index, sample_index)

# The recommended GLM-5.3-Flash caption is the Webshart default.
glm_caption = sample.caption

# Read the same sample's named fields when explicit selection is needed.
shard_name = dataset.get_shard_info(shard_index)["name"]
index_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=f"webshart/{shard_name}.json",
)
with open(index_path, encoding="utf-8") as handle:
    entry = json.load(handle)["files"][sample.path]

assert glm_caption == entry["caption"]
cogvlm_caption = entry["cogvlm_caption"]

SimpleTuner example

Add this object to your multidatabackend configuration:

{
  "id": "photo-aesthetics-webshart",
  "type": "webshart",
  "dataset_type": "image",
  "source": "webshart/terminusresearch-photo-aesthetics",
  "subfolder": "webshart",
  "caption_strategy": "webshart",
  "metadata_backend": "webshart",
  "webshart": {
    "cache_dir": "cache/webshart/photo-aesthetics",
    "shard_cache_gb": 25,
    "parallel_downloads": 4
  }
}

Set shard_cache_gb to 0 to use indexed HTTP range reads without whole-shard caching.

This configuration uses the GLM-5.3-Flash caption field. Use the explicit metadata loading example above if you need the legacy cogvlm_caption field instead.

Webshart Python example

import webshart

dataset = webshart.discover_dataset(
    "webshart/terminusresearch-photo-aesthetics",
    subfolder="webshart",
)
print(dataset.num_shards, dataset.total_files)
print(dataset.probe_caption_layout(max_shards=8))

loader = webshart.TarDataLoader(dataset)
sample_ref = dataset.list_samples_in_shard(0)[0]
sample = loader.load_sample(0, 0)
assert sample.path == sample_ref
print(sample.path, sample.caption, sample.width, sample.height)