gpt-j-6B: benchmark results
EleutherAI's GPT-J-6B, a 28-layer transformer trained with Mesh Transformer JAX using rotary position embeddings, sharing GPT-2/GPT-3's tokenizer, trained on English-only text. Provider: EleutherAI. Released 2021-06-09. Access: Open.
Aggregate Rating 1167 ± 11, rank #1609 of 1656 rated models, from 132 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| TIER (Qwen2.5-7B-Instruct Judge) - Contextual | 11.1 | Attack success rate (%; L3 harmful intent embedded in plausi | 100 |
| TIER (Qwen2.5-7B-Instruct Judge) - Euphemistic | 19.6 | Attack success rate (%; L2 euphemistic rewrites of the expli | 100 |
| HELM Classic - BLiMP | 83.4 | Exact Match (%) | 90.3 |
| MMLU-by-task - Machine Learning | 38.39 | Accuracy (%) | 75.5 |
| HELM Classic - Synthetic Reasoning Natural | 19.88 | F1 (%) | 60.3 |
| TIER (Llama-3.1-8B-Instruct Judge) - Contextual | 6.4 | Attack success rate (%; L3 harmful intent embedded in plausi | 60 |
| TIER (Llama-3.1-8B-Instruct Judge) - Jailbreak | 13.9 | Attack success rate (%; L4 jailbreak prompts: in-the-wild ja | 60 |
| TIER (Qwen2.5-7B-Instruct Judge) | 23.9 | Overall attack success rate (%; mean of the four threat leve | 60 |
| TIER (Qwen2.5-7B-Instruct Judge) - Jailbreak | 28 | Attack success rate (%; L4 jailbreak prompts: in-the-wild ja | 60 |
| HELM Classic - MATH | 11.08 | Equivalent (%) | 57.4 |
| HELM Classic - IMDB | 93.87 | Exact Match (%) | 56.1 |
| RealToxicityPrompts | 0.05 | Toxic fraction | 56.1 |
Interactive version: theaggregate.ai/model?slug=gpt-j-6b · How It Works · Data refreshed daily, snapshot 2026-10-20.