--- pretty_name: Adcom Captioned AI Images configs: - config_name: default data_files: - split: train path: captions.parquet tags: - images - captions - ai-generated-images - natural-language license: apache-2.0 language: - en --- # ADCOM Captioned AI Images ## Dataset description Adcom's captioned AI Images is a collection of images paired with associated captioning and descriptive metadata. A notebook is provided for sampling this dataset at https://huggingface.co/datasets/codeShare/adcom_captioned_ai_images/blob/main/sample_multi_author_dataset_fixed.ipynb This dataset does **not** have NSFW content aside from suggestive female oriented themes. The repository uses an indexed storage layout designed to make image retrieval more efficient. Instead of storing all image data directly inside a single metadata file, it separates the metadata index from the image data. The dataset consists of: - A Parquet metadata index: `captions.parquet` - JSONL image shards containing the image data - Byte offsets and record lengths in the Parquet index for direct image access This layout allows users to retrieve individual images without downloading and scanning every image shard. ## Repository structure ```text adcom_captioned_ai_images/ │ ├── README.md ├── captions.parquet │ └── [JSONL image shards] ├── ... └── ... ``` The exact shard paths are stored in the `json_file` column of `captions.parquet`. Use those paths rather than assuming fixed shard names. ### 1. Metadata index: `captions.parquet` The Parquet file contains one metadata/index row per image. | Column | Description | |---|---| | `idx` | Image identifier | | `prompt` | Text prompt or caption associated with the image | | `tags` | Descriptive tags associated with the image | | `author` | Author or contributor | | `json_file` | Repository-relative path to the JSONL shard containing the image | | `byte_offset` | Zero-based byte offset of the JSONL record within that shard | | `byte_length` | Number of bytes to read for that record | The index columns `json_file`, `byte_offset`, and `byte_length` are used together to locate an image record directly. Offsets and lengths refer to bytes, not character positions. ### 2. Image data: JSONL shards Images are stored in JSONL files. Each line represents an individual image record encoded as JSON. A record has the following structure: ```json { "idx": "IMAGE_ID", "image": "data:image/png;base64,..." } ``` The `image` field contains a base64-encoded image, commonly represented as a data URI. Consumers should also allow for a raw base64 string. The `idx` field identifies the image and corresponds to the `idx` value in the Parquet index. ### 3. How indexed retrieval works For an image with a given `idx`: 1. Find its row in `captions.parquet`. 2. Read the row's `json_file`, `byte_offset`, and `byte_length`. 3. Download the referenced JSONL shard. 4. Seek to `byte_offset` in binary mode. 5. Read `byte_length` bytes. 6. Parse the JSON record and decode its base64 image. This avoids scanning the entire shard for the selected image. --- ## Quick start ### Install dependencies ```bash pip install pandas pyarrow pillow huggingface_hub ``` ### Load the metadata index ```python import pandas as pd from huggingface_hub import hf_hub_download REPO_ID = "codeShare/adcom_captioned_ai_images" parquet_path = hf_hub_download( repo_id=REPO_ID, filename="captions.parquet", repo_type="dataset" ) metadata = pd.read_parquet(parquet_path) print(metadata.head()) print(f"Total indexed images: {len(metadata):,}") ``` For a private or gated repository, authenticate with a Hugging Face access token and pass it as the `token` argument to `hf_hub_download`. ### Retrieve one image by ID This example downloads only the shard containing the requested image and reads only the indexed record. ```python import json import base64 import re import io import pandas as pd from PIL import Image from huggingface_hub import hf_hub_download REPO_ID = "codeShare/adcom_captioned_ai_images" # Load the metadata index. parquet_path = hf_hub_download( repo_id=REPO_ID, filename="captions.parquet", repo_type="dataset" ) metadata = pd.read_parquet(parquet_path) # Choose an image ID from the index. target_idx = str(metadata.iloc[0]["idx"]) # Find its index entry. row = metadata.loc[ metadata["idx"].astype(str) == target_idx ].iloc[0] # Download only the required shard. shard_path = hf_hub_download( repo_id=REPO_ID, filename=str(row["json_file"]), repo_type="dataset" ) # Read the exact JSONL record using its byte location. with open(shard_path, "rb") as f: f.seek(int(row["byte_offset"])) raw_record = f.read(int(row["byte_length"])) record = json.loads(raw_record.decode("utf-8")) # Verify the retrieved record. assert str(record["idx"]) == target_idx # Decode the image. image_value = record["image"].strip() if image_value.startswith("data:") and "," in image_value: image_value = image_value.split(",", 1)[1] image_value = re.sub(r"\s+", "", image_value) image_bytes = base64.b64decode(image_value) image = Image.open(io.BytesIO(image_bytes)).convert("RGB") display(image) print("Image ID:", target_idx) print("Prompt:", row["prompt"]) print("Author:", row["author"]) print("Tags:", row["tags"]) ``` ### Retrieve a random sample of images The following example samples metadata first and downloads only the shards needed for the selected images. ```python import json import base64 import re import io import pandas as pd from PIL import Image from huggingface_hub import hf_hub_download REPO_ID = "codeShare/adcom_captioned_ai_images" NUM_SAMPLES = 10 RANDOM_SEED = 42 # Load the index. parquet_path = hf_hub_download( repo_id=REPO_ID, filename="captions.parquet", repo_type="dataset" ) metadata = pd.read_parquet(parquet_path) # Optional author filter. # Set to None to include all authors. AUTHOR_FILTER = "Adcom" if AUTHOR_FILTER is not None: metadata = metadata[ metadata["author"].fillna("").astype(str).str.casefold() == AUTHOR_FILTER.casefold() ] if metadata.empty: raise ValueError("No matching records found.") # Reproducible random sample. sampled = metadata.sample( n=min(NUM_SAMPLES, len(metadata)), random_state=RANDOM_SEED ) # Download only the shards required by the sample. shard_cache = {} for _, row in sampled.iterrows(): shard_name = str(row["json_file"]) if shard_name not in shard_cache: shard_cache[shard_name] = hf_hub_download( repo_id=REPO_ID, filename=shard_name, repo_type="dataset" ) # Retrieve the indexed JSONL record. with open(shard_cache[shard_name], "rb") as f: f.seek(int(row["byte_offset"])) raw_record = f.read(int(row["byte_length"])) record = json.loads(raw_record.decode("utf-8")) if str(record["idx"]) != str(row["idx"]): raise ValueError("Image ID does not match the metadata index.") # Decode image data. image_value = record["image"].strip() if image_value.startswith("data:") and "," in image_value: image_value = image_value.split(",", 1)[1] image_value = re.sub(r"\s+", "", image_value) image_bytes = base64.b64decode(image_value) image = Image.open(io.BytesIO(image_bytes)).convert("RGB") print("Image ID:", row["idx"]) print("Prompt:", row["prompt"]) print("Author:", row["author"]) display(image) image.close() ``` --- ## Performance notes - **Metadata-first sampling:** Select records from the Parquet index before downloading image data. - **Selective shard downloads:** Only download shards referenced by the selected records. - **Direct record access:** Use byte offsets and lengths instead of scanning JSONL files. - **Reproducible sampling:** Set `random_state` to a fixed integer to reproduce a sample. - **Private repository support:** Pass a Hugging Face token to the download functions when authentication is required. ## Important notes - The Parquet index and JSONL shards must remain consistent. If shard contents are rewritten or reordered, the corresponding byte offsets and lengths must be regenerated. - Image data is stored as base64 in JSONL records; it is decoded into image bytes when retrieved. - The dataset card does not declare a license. Please consult the dataset owner for licensing and usage permissions before redistributing or using the data commercially. ## Browsing images and captions in the repaired dataset This repository stores its caption index and image records separately: ```text captions.parquet images/ part_00000.jsonl part_00001.jsonl ... ``` ### How the files relate `captions.parquet` is the index. Each row describes a caption/image record and contains the information needed to locate the corresponding record in a JSONL shard. The important fields are: | Field | Meaning | |---|---| | `idx` | Global, sequential record ID across the dataset | | `author` | Author/source associated with the record | | `json_file` | Repository-relative path to the JSONL shard | | `byte_offset` | Starting byte position of the JSONL record | | `byte_length` | Number of bytes belonging to that record | The `json_file`, `byte_offset`, and `byte_length` fields work together: open the named JSONL shard in binary mode, seek to `byte_offset`, and read exactly `byte_length` bytes. Decode those bytes as UTF-8 and parse the resulting JSON. **Important:** `idx` is a global dataset ID. It is not an author-local ID and should not be interpreted as a line number or byte position. Use the byte reference fields to retrieve the JSONL record. The JSONL record contains the corresponding record data, including its global `idx` and author information. The Parquet row and JSONL record refer to the same logical item. ### 1. Install dependencies ```python !pip -q install -U huggingface_hub pyarrow pandas ``` ### 2. Load the Parquet index ```python import pyarrow.parquet as pq from huggingface_hub import hf_hub_download REPO_ID = "codeShare/adcom_captioned_ai_images" parquet_path = hf_hub_download( repo_id=REPO_ID, repo_type="dataset", filename="captions.parquet", ) captions = pq.read_table(parquet_path) print("Rows:", captions.num_rows) print("Columns:", captions.column_names) ``` ### 3. Find a record using its global `idx` For example, retrieve global record ID `123`: ```python import pyarrow.compute as pc def find_caption_row(global_idx): matches = pc.equal(captions["idx"], global_idx) result = captions.filter(matches) if result.num_rows == 0: raise KeyError(f"Global idx {global_idx} was not found.") if result.num_rows > 1: raise ValueError( f"Global idx {global_idx} occurs more than once." ) return result.to_pylist()[0] row = find_caption_row(123) print("Global idx:", row["idx"]) print("Author:", row["author"]) print("JSONL shard:", row["json_file"]) print("Byte offset:", row["byte_offset"]) print("Byte length:", row["byte_length"]) ``` You can also browse by author or other caption metadata using the Parquet columns: ```python import pandas as pd df = captions.to_pandas() print(df.head()) print(df["author"].value_counts()) ``` ### 4. Retrieve the exact JSONL record using its byte reference This is the recommended random-access method. It does not require reading the entire JSONL shard or counting lines. ```python import json def load_jsonl_record(row): shard_path = hf_hub_download( repo_id=REPO_ID, repo_type="dataset", filename=row["json_file"], ) offset = int(row["byte_offset"]) length = int(row["byte_length"]) with open(shard_path, "rb") as f: f.seek(offset) record_bytes = f.read(length) if len(record_bytes) != length: raise IOError( f"Expected {length} bytes, got {len(record_bytes)}." ) record = json.loads(record_bytes.decode("utf-8")) # Verify the Parquet-to-JSONL relationship. assert int(record["idx"]) == int(row["idx"]), ( f"ID mismatch: Parquet={row['idx']}, " f"JSONL={record.get('idx')}" ) if "author" in record and "author" in row: assert record["author"] == row["author"], ( "Author mismatch between Parquet and JSONL." ) return record record = load_jsonl_record(row) print(json.dumps(record, indent=2, ensure_ascii=False)) ``` ### 5. Browse a record by global ID Combine the two operations into one helper: ```python def get_record(global_idx): row = find_caption_row(global_idx) record = load_jsonl_record(row) return row, record row, record = get_record(123) print("Global ID:", row["idx"]) print("Author:", row["author"]) print("Shard:", row["json_file"]) print("Record:") print(json.dumps(record, indent=2, ensure_ascii=False)) ``` ### 6. Browse several records ```python global_ids = [0, 1, 2, 100, 123] for global_idx in global_ids: try: row, record = get_record(global_idx) print( f"\nID={global_idx} | " f"author={row['author']} | " f"shard={row['json_file']}" ) print(json.dumps(record, indent=2, ensure_ascii=False)) except KeyError: print(f"Global ID {global_idx} not found.") ``` ### 7. Inspect image-related fields The JSONL record is the source of the image data and associated metadata. Field names and image encoding should be inspected from the record rather than assumed. ```python row, record = get_record(123) print("Available fields:") for key, value in record.items(): print(f"{key}: {type(value).__name__}") print("\nPotential image-related fields:") for key, value in record.items(): if any(term in key.lower() for term in ("image", "img", "url", "path")): print(f"{key}: {str(value)[:300]}") ``` If a record contains an image URL or a base64 data URI, the following example can preview it: ```python import base64 import requests from io import BytesIO from PIL import Image from IPython.display import display def preview_image_field(record): # Common URL field names. for key in ("image_url", "url"): value = record.get(key) if isinstance(value, str) and value.startswith(("http://", "https://")): response = requests.get(value, timeout=30) response.raise_for_status() display(Image.open(BytesIO(response.content))) return # Common base64 data URI field. for key in ("image", "image_data", "image_base64"): value = record.get(key) if not isinstance(value, str): continue if value.startswith("data:image/") and "," in value: encoded = value.split(",", 1)[1] image_bytes = base64.b64decode(encoded) display(Image.open(BytesIO(image_bytes))) return if key == "image_base64": image_bytes = base64.b64decode(value) display(Image.open(BytesIO(image_bytes))) return print( "No recognized URL or base64 image field was found. " "Inspect the record fields above to determine its image encoding." ) row, record = get_record(123) preview_image_field(record) ``` ### Data integrity and access rules - Use `idx` as the global record identifier. - Use `json_file` to identify the JSONL shard. - Use `byte_offset` and `byte_length` together for exact random access. - Read JSONL files in binary mode when using byte offsets. - Decode the selected byte range as UTF-8 before parsing JSON. - Do not treat byte offsets as character offsets or line numbers. - Do not derive JSONL paths from author names or local IDs. - The Parquet index and JSONL record should agree on global `idx` and author. - When adding records, assign new global IDs after the current maximum and calculate byte offsets and lengths from the exact serialized UTF-8 JSONL bytes. This structure allows efficient lookup of an individual record without loading every JSONL shard into memory.