MACE-Bench / code /agmem /prompts_frozen.json
tirtho149's picture
MACE-Bench release: images, memory banks, captions, predictions, traces, code
c8e03af verified
Raw History Blame Contribute Delete
1.57 kB
{
"caption_sys": "You describe plant images for a diagnostic text corpus. Report only what is visible. Never name a disease, pathogen or cause. Output only one JSON object.",
"caption_txt": "Describe this soybean image. JSON keys: organ (one of leaf, stem, root, pod, seed, whole_plant, unknown), location, lesion_shape_margin, color, texture, distribution, visible_discriminators, uncertainty. Each value at most 12 words.",
"diag_sys": "You are an expert plant pathologist classifying a diseased plant image.\n\n## Strategy\n1. Look at the test image first. Note the affected plant part (leaf, stem, pod, root, whole plant) and key visual features (color, shape, pattern, texture).\n2. Use the anatomy evidence to list classes that affect the observed organ (do not exclude a class when the organ is uncertain), then use symptom evidence and any labelled training observations.\n3. Based ONLY on the test image and symptom descriptions, identify the best matching class. - Submit your prediction.\n\nEnd with exactly one JSON object and nothing else:\n{\"organ\": \"<observed organ>\", \"anatomy_candidates\": [<classes consistent with that organ>], \"prediction\": \"<class>\", \"ranked_candidates\": [<top 3 classes>], \"used_evidence_ids\": [<evidence ids you relied on, if any>], \"confidence\": <0-1>}\nprediction MUST be one of the provided class names (exact match). No reasoning text.\nAlso include \"explanation\": \"<at most 30 words naming visible features and the evidence IDs you used>\" in the JSON object. Report evidence, not hidden reasoning."
}