```json { "action": "final", "content": "## Missing Values Analysis\n\nThe dataset has **13 columns with missing values** out of 69 total columns. The missing data ranges from less than 1% to nearly 100%.\n\n### High Missing Rate (>90%)\nThese columns are almost entirely empty and likely unusable for most analyses:\n- **Manual evaluation**: 99.54% missing (1,305 of 1,311 rows)\n- **Citations**: 98.70% missing (1,294 rows)\n- **Finetune compute (FLOP)**: 98.47% missing (1,291 rows)\n- **Hardware utilization**: 98.40% missing (1,290 rows)\n- **Batch size**: 98.25% missing (1,288 rows)\n- **Epochs**: 96.80% missing (1,269 rows)\n- **Hardware quantity**: 96.19% missing (1,261 rows)\n- **Training compute cost (2023 USD)**: 95.42% missing (1,251 rows)\n- **Training dataset size (datapoints)**: 91.38% missing (1,198 rows)\n\n### Moderate Missing Rate (50-90%)\n- **Training compute (FLOP)**: 79.63% missing (1,044 rows)\n- **Parameters**: 75.36% missing (988 rows)\n\nThese columns have substantial data but still require careful handling in analyses.\n\n### Low Missing Rate (<1%)\n- **stderr**: 0.99% missing (13 rows)\n- **started_at**: 0.31% missing (4 rows)\n\nThese columns are nearly complete and reliable for most analytical purposes.\n\n### Key Takeaway\nThe majority of missing data is concentrated in training infrastructure and compute-related fields (hardware, batch size, epochs, compute costs). This suggests these metrics may only be available for a subset of models in the benchmark dataset, likely those where detailed training information was publicly disclosed.", "charts": [ { "kind": "hbar", "x": "column_name", "series": ["missing_percentage"], "title": "Missing Data by Column" } ] } ```