Atlas

Models

← All models

Qwen2.5 7B Instruct

Alibaba · 2024-09-19 · 7.6B parameters

Qwen2.5-7B-Instruct is Alibaba's dense 7B-class instruction-tuned language model from the Qwen2.5 family. Qwen reports improved instruction following, coding, mathematics, and structured output, while maintaining support for more than 29 languages. Its shipped configuration is set to 32,768 tokens, while Qwen documents support up to 131,072 tokens via YaRN extrapolation.

Benchmark scores

BenchmarkScore
BALROG7.8
BBH (Open LLM Leaderboard v2)53.9
BigCodeBench Complete46.1
BigCodeBench Instruct37.6
BigCodeBench-Hard Complete16.2
BigCodeBench-Hard Instruct12.2
BRIDGE Medical (chain-of-thought)30.3
BRIDGE Medical (few-shot)41.6
BRIDGE Medical (zero-shot)31.3
BuzzBench31.4
Capability113.1
EQ-Bench v269.2
EQ-Bench v2 + MAGI-Hard Combined62.6
Global PIQA - Non-Parallel (Strict Exact Match)66.5
Global PIQA - Parallel (Strict Exact Match)42.1
GPQA (Open LLM Leaderboard v2)29.1
IFEval (Open LLM Leaderboard v2)75.9
LegalBench69.6
LegalBench - Conclusion Tasks65.6
LegalBench - Interpretation Tasks73.3
LegalBench - Issue Tasks72.6
LegalBench - Rhetoric Tasks73.0
LegalBench - Rule Tasks63.3
MAGI-Hard56.0
MATH Level 5 (Open LLM Leaderboard v2)50.0
MEDIC (clinical summarization)86.2
MEDIC (closed-ended)60.0
MEDIC (open-ended Elo)1372
MMLU72.9
MMLU-Pro (Open LLM Leaderboard v2)42.9
Loading Atlas data…