Every condition uses the same AG News test rows, vocabulary, PyTorch architecture, hyperparameters, and three fixed random seeds. No LLM call is made by this benchmark.
| Condition | Rows | Accuracy | Macro-F1 |
|---|---|---|---|
| original_dataset | 3000 | 0.8315 ± 0.0074 | 0.8307 ± 0.0076 |
| regressed_generated | 3000 | 0.8070 ± 0.0040 | 0.8061 ± 0.0037 |
| repaired_generated | 3000 | 0.8315 ± 0.0074 | 0.8307 ± 0.0076 |
The repaired generated data recovers +2.45 accuracy points and finishes +0.00 points from the matching original control. It changes 0 texts, compared with 2,767 in the regressed output.
The original condition contains all 3,000 source rows plus disclosed conflicting-label duplicates, class-skewing duplicates, and empty rows. BuffData generates the optimized condition from the exact same input.
| Condition | Rows | Accuracy | Macro-F1 |
|---|---|---|---|
| original_contaminated | 5700 | 0.7515 ± 0.0088 | 0.7509 ± 0.0077 |
| generated_optimized | 3000 | 0.8425 ± 0.0088 | 0.8415 ± 0.0089 |
Generated optimized data improves accuracy by +9.10 points and macro-F1 by +9.06 points.
- Regression recovery: PASS
- Strict original preservation (generated accuracy >= original and identical text): PASS
- Contaminated-data value: PASS
These results establish behavior for these fixed samples and this lightweight classifier; they do not promise the same gain on every dataset or model.