gpt-neox-20B: benchmark results
EleutherAI's GPT-NeoX-20B, a 20B autoregressive model trained on the Pile using the GPT-NeoX library, with an architecture resembling GPT-3 and nearly identical to GPT-J-6B. Provider: EleutherAI. Released 2022-02-02. Access: Open.
Aggregate Rating 1186 ± 11, rank #1600 of 1656 rated models, from 128 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM Classic - BLiMP | 83.85 | Exact Match (%) | 96.8 |
| HELM Classic - Dyck | 74.73 | Exact Match (%) | 92.6 |
| Open Korean LLM - Ko-GPQA-Diamond | 25.76 | Normalized accuracy (%) | 81.3 |
| HELM Classic - IMDB | 94.8 | Exact Match (%) | 72.7 |
| HELM Classic - MATH | 14.05 | Equivalent (%) | 72.1 |
| HELM Classic - MATH Chain-of-Thought | 7.06 | Equivalent (%) | 64.7 |
| RealToxicityPrompts | 0.05 | Toxic fraction | 61 |
| HELM Classic - Entity Matching | 82.03 | Exact Match (%) | 51.5 |
| HELM Classic - Synthetic Reasoning Natural | 16.73 | F1 (%) | 51.5 |
| HELM Classic - MS MARCO TREC | 39.82 | NDCG@10 (%) | 50 |
| HELM Classic - NaturalQuestions Open Book | 59.61 | F1 (%) | 47.7 |
| HELM Classic - GSM8K | 5.27 | Exact Match (%) | 47.1 |
Interactive version: theaggregate.ai/model?slug=gpt-neox-20b · How It Works · Data refreshed daily, snapshot 2026-10-20.