EmpathicRobotics/imagenet-variations-synth-pilot

ImageNet Variations Synth Pilot (1K)

Synthetic vision-language instruction pilot from laion/imagenet_variations prompts.

Pipeline

1. Sample 1000 prompts from imagenet_variations
2. Add one style framing per prompt (website / textbook / diagram / …)
3. Generate image with FLUX.1-schnell
4. Encode with SEED-2 (<seed2_N>, 32 tokens)
5. Qwen2.5-VL-7B-Instruct looks at the actual image and writes instruction + response
   (modes: describe / QA / howto / reverse)
6. Format: capitalized USER / ASSISTANT (synth-llava style)

Format

USER: <seed2> <seed2_N> … </seed2> {instruction} ASSISTANT: {response}

Each JSONL row also has image_path (relative under images/) and metadata (style, mode, base prompt, models).

Contents

Rows	1000
Images	1000 × 1024² PNG under images/
Failures	0

Files

• data/pilot.jsonl — records
• images/*.png — generated images (inv_XXXXXX.png)

Notes

Pilot for review before scaling. Intended for VLA / multimodal pretraining research.