Llama 3.1 8B Instruct: benchmark results
Meta Llama 3.1 8B instruction-tuned checkpoint. Provider: Meta. Released 2024-07-23. Access: Open.
Aggregate Rating 1442 ± 3, rank #1035 of 1656 rated models, from 3757 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BAGEL - Xeno-canto Harmonic Structure | 46.4 | Accuracy (%, times 100) on the 427 Xeno-canto questions on h | 100 |
| DropKV Reference - LongBench (full KV cache) | 49.91 | Average score | 100 |
| DropKV Reference - RULER 16K (full KV cache) | 93.63 | Accuracy (%) | 100 |
| DropKV Reference - RULER 32K (full KV cache) | 92.56 | Accuracy (%) | 100 |
| DropKV Reference - RULER 64K (full KV cache) | 91.35 | Accuracy (%) | 100 |
| FDARxBench - Refusal F1 | 79.6 | Macro F1 (x 100) for abstaining on unanswerable questions (a | 100 |
| HELM TORR - Fin Qa - Robustness | 63 | Robustness (0-100) | 100 |
| JailNewsBench (Role Play) - Harmfulness Score | 2.3 | Mean harmfulness of the generated fake news (0 to 4): the av | 100 |
| LIBRA - MatreshkaYesNo | 75 | Dataset Total Score (%) | 100 |
| MASLegalBench | 86.21 | Best accuracy (imported) | 100 |
| Open CoT - LSAT Reading Comprehension | 27.51 | CoT Gain (%) | 100 |
| PSI-Bench (Depression Simulation) - PATIENT-Psi | 61.8 | Overall fidelity score (%) of the model as a depressed-patie | 100 |
Interactive version: theaggregate.ai/model?slug=llama-3-1-8b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.