Gemma 7B.
Google DeepMind Released February 21st, 2024 Family Gemma 1 The lab's model card Compare with others Google DeepMind Feb 2024 Open weights Open model
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BoolQ | 83.2% | #11 of 28 | Lab model card | |
| Open Arabic LLM Leaderboard v1 | 46% | #23 of 55 | Official leaderboard | |
| MuSR | 40.6% | #55 of 98 | Official leaderboard | |
| SIQA | 51.8% | #14 of 24 | Lab model card | |
| OpenBookQA | 52.8% | #15 of 24 | Lab model card | |
| ANLI | 48.7% | #7 of 11 | Claim in Phi-3 Technical Report: A Highly Capable Language Model L…, via Epoch AI | |
| ARC-Easy | 81.5% | #15 of 23 | Lab model card | |
| PIQA | 81.2% | #25 of 37 | Lab model card | |
| HellaSwag | 81.2% | #38 of 53 | Lab model card | |
| EQ-Bench v2 | 61.7 | #64 of 85 | Official leaderboard | |
| WinoGrande | 72.3% | #36 of 47 | Lab model card | |
| Natural Questions | 23% | #16 of 20 | Lab model card | |
| MMLU-Pro | 33.7% | #174 of 212 | Official leaderboard | |
| AlpacaEval 2.0 | 10.4% | #39 of 47 | Official leaderboard | |
| CommonSenseQA | 71.3% | #16 of 19 | Lab model card | |
| BIG-Bench Hard | 55.1% | #62 of 73 | Lab model card | |
| AGIEval | 41.7% | #34 of 40 | Lab model card | |
| MMLU | 64.3% | #113 of 131 | Lab model card | |
| TriviaQA | 63.4% | #35 of 40 | Lab model card | |
| ARC-Challenge | 53.2% | #43 of 49 | Lab model card | |
| Vectara Hallucination Leaderboard (HHEM) | 14.8% | #138 of 148 | Official leaderboard | |
| LMArena (Elo) | 1,138 | #313 of 335 | Official leaderboard | |
| TruthfulQA | 31.8% | #29 of 30 | Lab model card | |
| WildBench | 10.2 | #46 of 47 | Official leaderboard | |
| TurkishMMLU | 43.6% | #8 of 8 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MBPP+ | 43.4% | #30 of 35 | Official leaderboard | |
| HumanEval+ | 28.7% | #40 of 46 | Official leaderboard | |
| MBPP | 44.4% | #65 of 71 | Lab model card | |
| HumanEval | 32.3% | #94 of 101 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SuperCLUE-Agent | 51.3 | #7 of 7 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MedQA | 51.3% | #60 of 71 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CrowS-Pairs | 51.3% | #1 of 4 | Lab model card | |
| OR-Bench | 26.3% | #9 of 28 | Official leaderboard | |
| BOLD | 49.1 | #1 of 2 | Lab model card | |
| Winogender | 54.2% | #3 of 4 | Lab model card | |
| CyberMetric | 34.3% | #8 of 10 | Official leaderboard | |
| ToxiGen | 39.6 | #4 of 4 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Open LLM Leaderboard v2 | 15.4 | #69 of 99 | Official leaderboard | |
| HELM Lite | 33.6 | #56 of 71 | Official leaderboard |
The Self-Reported Index fits one score to the 19 capability results Google DeepMind reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 97.3 to 101.5. How we compute it.
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
Google DeepMind's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 8 of them aren't in the launch post.