HomesteaderLabs/forager-sightings

Forager Sightings

A growing, crowd-contributed dataset of wild plant, mushroom, and berry photos with human labels, collected through the "Beat the Machine" game in Forager's Field Station.

Every row pairs a field photo with the contributor's own identification and the on-device model's prediction (or its refusal). It is an open-research dataset for improving small, on-device foraging models — especially the hard cases the model abstains on.

How it's collected

Players upload a photo and guess what it is. The same router→expert→abstain pipeline that runs on the handheld device makes its own call. Both are recorded. Only in-domain photos (berry / mushroom / plant) are stored — out-of-domain uploads are politely rejected and discarded.

Schema (metadata.jsonl)

field	type	meaning
file_name	str	path to the image under images/
user_label	str	the contributor's free-text / chosen identification
machine_prediction	str	species key the model committed to (or unknown)
machine_confidence	float	model confidence [0,1]
machine_abstained	bool	true if the model refused to commit
machine_safety	str	SAFE / CAUTION / DEADLY / UNKNOWN (model's tier)
routed_domain	str	berry / mushroom / plant
contributor	str	Hugging Face handle (via login)
consent	bool	always true — submission requires consent
license	str	CC-BY-4.0
timestamp	str	ISO-8601 UTC
phash	str	64-bit perceptual hash (dHash) for near-duplicate rejection

⚠️ Safety

Labels are self-reported and unverified — this is weakly-labeled data intended for research and curation, not a field authority. Never make a consumption decision from this dataset. Wild mushroom and plant identification carries fatal risk.

License

CC-BY-4.0 — free to use with attribution to HomesteaderLabs.