• Citrus Doctor's strict accuracy rose from 14.9% to 52.8% on 335 test photos, per its model card, but strict accuracy on healthy leaves is 0.0%.
• The 0.0% healthy score is mainly a strict name mismatch, because the model never repeats 'Healthy citrus tissue' verbatim. It is not simply a failure to see healthy leaves.
• The Agate 4-step student scored 0.536 GenEval against the teacher's 0.563 at 50 steps. Their 95% intervals (0.508-0.564 and 0.534-0.591) overlap substantially, so the gap may not be significant.
• The Agate card's only paired interval compares the student with the previous 4-step version (+0.027, 95% interval +0.014 to +0.041). It reports no paired interval against the teacher.
• The citrus card says its relaxed 61.2% metric uses an alias table tuned on the test set, so only the strict number is the primary result.

Hugging Face's post argues that an agent can build the model you could not find. The model cards support a narrower claim. Across six projects, ML Intern fine-tuned or distilled small models for about $103 in compute, according to the post. The two cards checked for this piece, Citrus Doctor and Agate Preview 002, disclose weak spots and statistical limits that the headline numbers leave out. The other four projects rest on the post alone.

The citrus gain is real but narrow

Citrus Doctor more than tripled the base model's strict accuracy, from 14.9% to 52.8% on 335 test photos, according to its model card. The same card reports 69.1% on bacterial disease and 23.7% on nutrient deficiency. Relaxed matching reaches 61.2% overall, but the card says its alias table was tuned on the test set, so the strict metric is the primary result. The card also advises that an agronomist review any treatment text.

A 0.0% healthy score is mostly a naming problem

The card reports 0.0% strict accuracy on healthy leaves, and it attributes this mainly to a strict name mismatch: the model never repeats 'Healthy citrus tissue' verbatim. A 'no symptoms' template appears in only 18 of 38 healthy-image answers. The practical lesson is that a strict string-match metric can hide whether the model recognises healthy tissue at all, while the relaxed metric that would show it was tuned on the test set.

The Agate gap to the teacher is not shown to be real

The 4-step Agate student scored 0.536 on GenEval against 0.563 for the teacher at 50 steps. The 95% intervals, 0.508 to 0.564 and 0.534 to 0.591, overlap substantially, so the card does not establish a quality loss. Its only paired interval, +0.027 with a 95% range of +0.014 to +0.041, compares the student with the previous 4-step version. The card reports no paired interval against the teacher.

The card still lists concrete weaknesses

Agate's card reports a FID-5k of 12.57 against 11.40 for the teacher, and about 12.8x faster generation, measured on the previous version with the same network and step count. It names exact text, counts above three and negation as weak. It says the gap to the teacher comes mostly from two-object and colour-credit prompts. It also says people can come out partly unclothed. These caveats are more useful to a reader than the single benchmark score.

Verdict: a measured first draft, not a production sign-off

A team with a narrow task, a dataset and a small budget can use this workflow to get a measured baseline quickly. But the cards show why a person must still read per-class results and confidence intervals: a 52.8% headline sits next to a 0.0% strict class, and a 0.027 benchmark gap may be within noise. Treat the agent's output as a draft of a training run. Stated costs and the other four projects remain unverified claims from the post.