Abstract
A Qwen3.5-4B model trained on measured Nutrition5k targets scored 85.2% on a frozen three-class calorie test. The best frontier model tested scored 75.8%. Label source changed downstream model quality, and removing 111 contested examples caused a larger loss than changing model size. A second phase extends the same 4B SFT stack to ten-class prediction, continuous regression, and ingredient-mass extraction. Those extensions show strong validation gains, but are not yet frontier comparisons.
The question
The claim is not that a 4B model is generally smarter. The question is whether general-purpose scale is still necessary after the task, labels, and hard cases are fixed.
The experiment uses Nutrition5k. Its calorie and ingredient targets come from physical measurements, so every model and every label source can be checked against the same answer.
What the inputs look like
Nutrition5k contains ordinary cafeteria plates rather than studio food photography. Each plate was captured from multiple side angles and from above, with measured dish mass, calories, macronutrients, and ingredient-level metadata.
Official Nutrition5k sample imagery · Google Research · CC BY 4.0
A 4B fine-tune beat every frontier model tested
The gold-label 4B fine-tune scored 85.2% on the frozen 822-case test. The strongest frontier model tested scored 75.8%. The same 4B model scored 50.6% before fine-tuning.
Frozen-test accuracy after supervised fine-tuning.
Same cases and the same output contract.
Frozen bucket accuracy
Measured ground truth · n=822
Figure 1. This is a task-fit result, not a general intelligence ranking.
The model inherited the quality of its labels
One qualified human read was 78.7% accurate. Weighted consensus from about 18 reads reached 86.9%. The trained model followed the same order: 78.5% from single reads, 80.3% from consensus, and 85.2% from measured truth.
Label quality → model quality
Label accuracy on the left; frozen-test accuracy after training on the right
Single human read
Weighted consensus
Figure 2. Aggregation improved the label source by 8.2 points; the gain survived model training.
The model did not erase label noise. It compressed the quality ceiling of its supervision into a cheap inference model.
The 111 contested cases mattered more than scale
Removing about 5% of the training set reduced full-test accuracy from 85.2% to 78.5%. On 42 unseen hard cases, accuracy fell from 71.4% to 33.3%.
Hard-case ablation
Same 4B recipe and budget; only the 111 contested training cases changed
Full frozen test
42 unseen hard cases
Figure 3. Targeted hard examples produced the largest measured change in the study.
The operational implication is direct: route labeling effort toward examples that break current models. Adding easy examples can increase volume without adding much information.
The same 4B stack learned three harder outputs
The post-2 suite changes the target while keeping the same base model and measured image population. These are validation results on 247 cases.
10-class bins
53.0%Exact accuracy · 0.59 mean bin error · 9.7% errors ≥2 bins
Continuous calories
47.5kcal MAE · 21.0 median error · 70.9% within ±50 kcal
Ingredient mass
75.0%Recall · 80.3% precision · 10.1 g mean absolute error
These new targets use measured gold labels. They extend the SFT method across output forms; they do not extend the human-consensus result or show a frontier-model win.
References
- Thames et al. Nutrition5k. CVPR, 2021.
- Hoffmann et al. Training Compute-Optimal Large Language Models. 2022.
- Swayamdipta et al. Dataset Cartography. EMNLP, 2020.
- Dawid and Skene. Maximum Likelihood Estimation of Observer Error-Rates Using the EM Algorithm. 1979.