Eval Cogito V2 Preview 70B: Large Language Model Benchmark Results

NousResearch's evaluation results for the Cogito v2 preview 70B model across multiple reasoning benchmarks, published on Hugging Face in August 2025. The dataset includes scores for tasks like math (AIME), creative writing, and general knowledge (GPQA). Metrics such as win rates and pass rates are provided for each benchmark.

Use Cases

• Compare model performance on mathematical reasoning tasks based on AIME benchmark scores.
• Evaluate creative writing quality based on the creative_writing_score metric.
• Assess general knowledge and problem-solving abilities using the GPQA and BBH benchmark results.
• Analyze model consistency and overlong response rates across different evaluation categories.

Strengths

• Includes results from at least 8 distinct benchmarks, such as AIME, ArenaHard, and DROP.
• Provides specific performance metrics like win rates (e.g., 86.9% on ArenaHard) and pass rates.
• Reports overlong response rates for several benchmarks, offering insight into model behavior.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is a preview version, and the full description is hosted externally.

Provenance

Source

NousResearch via Hugging Face.

Collection Method

Likely contains aggregated results from automated evaluation runs of the Cogito v2 preview 70B model.

Freshness

Last updated 2025-08-28 08:35:58; freshness should be verified.