carl-33B: benchmark results
An experimental conversation fine-tune released by ajibawa-2023, trained on over 100,000 synthetic therapy-style conversation sets. The card explicitly says it should not be used as a human therapist. Provider: ajibawa-2023. Access: Open.
Aggregate Rating 1360 ± 15, rank #7431 of 10916 rated models, from 68 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MMLU-by-task - ARC Challenge | 63.14 | Accuracy (%) | 93 |
| MMLU-by-task - College Mathematics | 39 | Accuracy (%) | 92.7 |
| MMLU-by-task - Marketing | 87.18 | Accuracy (%) | 91.8 |
| MMLU-by-task - HellaSwag | 66.35 | Accuracy (%) | 91.4 |
| MMLU-by-task - Formal Logic | 43.65 | Accuracy (%) | 91.3 |
| MMLU-by-task - Machine Learning | 47.32 | Accuracy (%) | 91.1 |
| MMLU-by-task - College Physics | 36.27 | Accuracy (%) | 90.7 |
| MMLU-by-task - College Medicine | 57.23 | Accuracy (%) | 90.3 |
| MMLU-by-task - Public Relations | 69.09 | Accuracy (%) | 90.2 |
| MMLU-by-task - US Foreign Policy | 85 | Accuracy (%) | 89.3 |
| MMLU-by-task - Virology | 50.6 | Accuracy (%) | 89.1 |
| MMLU-by-task - Conceptual Physics | 51.49 | Accuracy (%) | 89 |
Interactive version: theaggregate.ai/model?slug=carl-33b · How It Works · Data refreshed daily, snapshot 2026-11-08.