UPShf/FlowTalk-V1.1_ImageNet-1k-captions_captions-only

UPShf/FlowTalk (planned) V1.1 - ImageNet-1k captions (captions only)

This repository is intended to be the captions-only release used for UPShf/FlowTalk (planned) V1.1 training.

It does not ship any ImageNet images (public redistribution is typically not permitted). Users must link these captions to their own local copy of ImageNet-1k images (for example the 256x256 variant from benjamin-paine/imagenet-1k-256x256).

Base images: benjamin-paine/imagenet-1k-256x256.

Caption generation pipeline: Sigma-Captioner (SGLANG branch): https://github.com/uninterruptedpowersupply3-NEW/Sigma-Captioner/tree/SGLANG

What is in the dataset

This repo can contain two captions-only JSONL formats (both are sharded, and may also be provided as a best store-level (in WinRar) ZIP with "Best" compression):

A) Path-mapped captions (recommended)

Files:

• mapped-*.jsonl (keyed by ImageNet parquet image.path)

Each JSONL line looks like:

{"imagenet_path":"...","split":"train|validation|test","label":123,"global_idx":456789,"caption":"...","caption_source":"..."}

Notes:

• imagenet_path is the stable identifier from the parquet column image.path in benjamin-paine/imagenet-1k-256x256.
• Some rows are intentionally missing (low quality filtering / evaluation removals), so you should expect a partial match vs the full ImageNet-1k set.

B) pHash captions (robust to renames / slight re-encodes)

Files:

• captions-*.jsonl (keyed by a perceptual hash stored in the sha256 field for historical reasons)

Each JSONL line looks like:

{"sha256":"d8fa381417597327","hash_mode":"perceptual_hash_phash","caption":"...","caption_source":"...","image_filename":"..."}

Notes:

• hash_mode=perceptual_hash_phash means sha256 is NOT a real SHA-256; it is a pHash string (16 hex chars).

Build stats (current release)

• Captions exported / mapped: 891,549
• JSONs with no caption (filtered out): 2,298
• Full ImageNet-1k rows scanned (parquet): 1,431,167

Linking captions to your local images (order does not matter)

If your images are extracted in a different order (e.g. "image 12 is image 31"), you can still link captions by filename:

• Find the local file that matches imagenet_path.
• Use split to know whether it is train/validation/test.

If your local copy renamed files, use the pHash JSONL:

• Compute the same pHash (imagehash.phash) over your local images
• Join on the 16-hex string in the sha256 field

Warning:

• pHash is designed for similarity, not cryptographic uniqueness. If you want extra safety, do a second-stage verification (e.g., also compare byte-sha256 after a candidate match).

How to reproduce locally (Windows)

From the root of this workspace, this command:

• uses both caption sources (CaptinedIMGNET and extracted_images)
• reads only parquet metadata (image.path, label) and caption JSONs
• does not load image bytes

python .\codexGPT5.2HIGH.py build-mapped-index `
  --parquet_dir .\imagenet-1k-256x256\data `
  --out_dir .\flowtalk_mapped_shards `
  --out_zip .\flowtalk_v1.1_imagenet1k_captions_mapped_store.zip `
  --batch_size 50000 --queue 50000 --shard_size 50000 --workers 8

To be explicit (no auto-detect), add:

• --pairs_dir .\imagenet-1k-256x256\CaptinedIMGNET
• --pairs_dir .\imagenet-1k-256x256\extracted_images

To generate the pHash JSONL:

python .\codexGPT5.2HIGH.py export-captions `
  --caption_dir .\imagenet-1k-256x256\CaptinedIMGNET `
  --caption_dir .\imagenet-1k-256x256\extracted_images `
  --out_dir .\flowtalk_phash_shards `
  --hash_mode phash `
  --shard_size 50000 --queue 5000 --workers 8 `
  --skip_missing_images

Performance notes

• --pairs_lookup ram (default) scans the caption directories once and does O(1) in-RAM lookups instead of millions of per-row exists() calls.
• Optional (faster JSON): pip install orjson (the script uses it automatically when available).
• Optional (progress bar): pip install tqdm.
• Optional (perceptual hashing for export-captions --hash_mode phash): pip install ImageHash (imports as imagehash).

Licensing & Dataset Usage

Disclaimer: This section is provided for informational purposes only and does not constitute legal advice.

• Captions & Metadata: The text captions and JSON metadata generated in this repository are released under the Apache-2.0 license.
• Underlying Images: The original ImageNet images are NOT redistributed in this repository. Users must obtain the images independently and comply with the official ImageNet terms of access, as well as any upstream dataset terms (such as benjamin-paine/imagenet-1k-256x256).
• Copyright: This dataset provides derivative text descriptions. Users are responsible for ensuring their use of the combined text and image data complies with all applicable licenses.

Known Limitations & Bugs

• Language & Vocabulary Constraints: This dataset is intended to be entirely in English. However, because the captions and tags were generated by automated AI models, there are a few edge cases to be aware of:
  • Hallucinations: Rare instances of non-English characters or words may occur due to standard Vision-Language Model hallucinations.
  • Loanwords & Entities: Tag-based captions may include widely accepted loanwords (e.g., "taco", "sushi"), proper nouns, or domain-specific terminology that some strict language filters might flag as non-English.

If you are training a strict English-only model, you may want to apply a basic vocabulary filter to the text before training to catch any edge cases.

Model credits (captioning)

Captions were produced using a multi-model pipeline, including:

• BLIP captions: Salesforce/blip-image-captioning-large (stored as blip.caption)
• Tag-style captions (previously labeled wd_tagger in JSON): Qwen/Qwen3-VL-Embedding-2B using UPShf/Vocabulary-Qwen3-VL-Embedding-2B (stored as wd_tagger.caption)
• Sigma-Captioner (SGLANG) QA captions: Qwen/Qwen3.5-2B (stored under sglang.qa_pairs in the raw JSON)

Captioner mix (from a metadata scan over 893,847 caption JSON files; selection order sglang > blip > wd_tagger):

Captioner	Count	%
Salesforce/blip-image-captioning-large	496,043	55.50%
Qwen/Qwen3-VL-Embedding-2B (+ UPShf/Vocabulary-Qwen3-VL-Embedding-2B)	376,544	42.13%
Qwen/Qwen3.5-2B (Sigma-Captioner / SGLANG)	21,260	2.38%