| # Model comparison — sentiment | |
| Test split identical across versions (fingerprint a9ff6039d984, n=51). | |
| | | sentiment@v1 | sentiment@v2 | | |
| |---|---|---| | |
| | tier | encoder | encoder | | |
| | base | `sentence-transformers/all-MiniLM-L6-v2` | `sentence-transformers/all-MiniLM-L6-v2` | | |
| | **macro F1** | **0.882** | **0.881** | | |
| | accuracy | 88.2% | 88.2% | | |
| | coverage @ threshold | 82.4% (95.2% accurate) | 51.0% (96.2% accurate) | | |
| | escalation rate | 17.6% | 49.0% | | |
| | ECE after calibration | 0.066 | 0.093 | | |
| | download | 23.7 MB | 24.31 MB | | |
| | latency p95, python | 3.99 ms (cpu) | 3.57 ms (cpu) | | |
| | latency p95, browser (best) | — | — | | |
| | train time | 19 s | 120 s | | |
| | F1 · positive | 0.875 | 0.919 | | |
| | F1 · neutral | 0.914 | 0.875 | | |
| | F1 · negative | 0.857 | 0.848 | | |
| Prediction agreement on the test split: sentiment@v1 vs sentiment@v2: 88.2% | |
| Targets: macro F1 ≥ 0.9, download ≤ 30.0 MB. | |