codegen-16B-nl: benchmark results
Salesforce's CodeGen-NL 16B, an autoregressive program-synthesis model pretrained on the Pile dataset, the largest of four CodeGen sizes (350M, 2B, 6B, 16B) in the NL pre-training variant. Provider: Salesforce. Released 2022-04-13. Access: Open.
Aggregate Rating 1203 ± 12, rank #1579 of 1656 rated models, from 78 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MMLU-by-task - Moral Scenarios | 28.27 | Accuracy (%) | 63.2 |
| ToolBench - Tabletop | 16.2 | Task Score | 57 |
| MMLU-by-task - Machine Learning | 33.04 | Accuracy (%) | 53.4 |
| MMLU-by-task - College Physics | 25.49 | Accuracy (%) | 53.2 |
| MMLU-by-task - Abstract Algebra | 30 | Accuracy (%) | 52.2 |
| ToolBench - VirtualHome | 18 | Task Score | 51.2 |
| MMLU-by-task - High School Mathematics | 27.04 | Accuracy (%) | 50.6 |
| MMLU-by-task - Electrical Engineering | 42.76 | Accuracy (%) | 49.4 |
| MMLU-by-task - High School Physics | 29.8 | Accuracy (%) | 46.5 |
| ToolBench - The Cat API | 75 | Task Score | 45.3 |
| ToolBench - WebShop Long | 0 | Task Score | 44.2 |
| MMLU-by-task - Security Studies | 42.86 | Accuracy (%) | 43.3 |
Interactive version: theaggregate.ai/model?slug=codegen-16b-nl · How It Works · Data refreshed daily, snapshot 2026-10-20.