UNA-SimpleSmaug-34B-v1beta: benchmark results
fblgit's UNA fine-tune of Smaug-34B on the simple-math dataset, applying UNA to attention layers to boost math and reasoning. Provider: fblgit. Released 2024-02-05. Access: Open.
Aggregate Rating 1521 ± 23, rank #712 of 1656 rated models, from 25 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - ARC Challenge | 74.57 | Normalized accuracy (%) (25-shot) | 99.3 |
| Open Chinese LLM - WinoGrande | 72.14 | Accuracy (%) | 99.1 |
| Open LLM Leaderboard v1 - GSM8K | 72.48 | Accuracy (%) (5-shot) | 97.9 |
| Open LLM Leaderboard v1 - MMLU | 76.68 | Accuracy (%) (5-shot) | 97.9 |
| Open Chinese LLM - ARC Challenge | 66.04 | Accuracy (%) | 97.3 |
| Open Chinese LLM Leaderboard | 70.7 | Official Nonzero-Task Mean (%) | 96.4 |
| Open Chinese LLM - C-Eval Semantic | 90.6 | Accuracy (%) | 95.9 |
| Open Chinese LLM - HellaSwag | 71.08 | Accuracy (%) | 95.4 |
| Open Portuguese LLM - FaQuAD NLI | 80.98 | Macro F1 (%) | 91.4 |
| Open LLM Leaderboard v1 - WinoGrande | 83.82 | Accuracy (%) (5-shot) | 90.8 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 70.17 | MC2 (%) (0-shot) | 90 |
| Open Portuguese LLM - BLUEX | 65.37 | Accuracy (%) | 86.5 |
Interactive version: theaggregate.ai/model?slug=una-simplesmaug-34b-v1beta · How It Works · Data refreshed daily, snapshot 2026-10-20.