google

/

gemma-3-4b-it

like 1.02k

Follow

Google 37.8k

Added structured evaluation results from README benchmark tables covering 4 categories:

1. Reasoning & Factuality (11 benchmarks):

• HellaSwag: 77.2, BoolQ: 72.3, PIQA: 79.6, SocialIQA: 51.9
• TriviaQA: 65.8, Natural Questions: 20.0
• ARC-c: 56.2, ARC-e: 82.4, WinoGrande: 64.7
• BIG-Bench Hard: 50.9, DROP: 60.1

2. STEM & Code (8 benchmarks):

• MMLU: 59.6, MMLU Pro COT: 29.2, AGIEval: 42.1
• MATH: 24.2, GSM8K: 38.4, GPQA: 15.0
• MBPP: 46.0, HumanEval: 36.0

3. Multilingual (7 benchmarks):

• MGSM: 34.7, Global-MMLU-Lite: 57.0
• WMT24++ (ChrF): 48.4, FloRes: 39.2, XQuAD: 68.0
• ECLeKTic: 11.0, IndicGenBench: 57.2

4. Multimodal (15 benchmarks):

• COCOcap: 102.0, DocVQA: 72.8, InfoVQA: 44.1, MMMU: 39.2
• TextVQA: 58.9, RealWorldQA: 45.5, ReMI: 27.3
• AI2D: 63.2, ChartQA: 63.6, VQAv2: 63.9
• BLINK: 38.0, OKVQA: 51.0, TallyQA: 42.5
• SpatialSense VQA: 50.9, CountBenchQA: 26.1

Total: 41 benchmarks across reasoning, STEM, code, multilingual, and multimodal capabilities.

This enables the model to appear in leaderboards and makes it easier to compare with other models.

Note: Existing PRs (#57, #49, #34) modify README text content. This PR adds structured metadata to the YAML frontmatter and should not conflict.