GPT-NeoX-20B
EleutherAI · 2022-04-07 · 20.6B parameters
GPT-NeoX-20B is EleutherAI's dense 20,554,567,680-parameter autoregressive language model trained on the Pile, with weights released under Apache-2.0 in April 2022. It uses 44 transformer layers, a 2,048-token sequence length, rotary positional embeddings on 25% of the embedding dimensions, and parallel attention and feed-forward blocks. At paper submission, the authors described it as the largest dense autoregressive model with publicly available weights.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC AI2 | 41.1 |
| BBH (Open LLM Leaderboard v2) | 31.7 |
| Capability | 78.2 |
| GPQA (Open LLM Leaderboard v2) | 24.3 |
| GSM8K | 5.3 |
| IFEval (Open LLM Leaderboard v2) | 25.9 |
| MATH Level 5 (Open LLM Leaderboard v2) | 1.4 |
| MMLU | 33.6 |
| MMLU-Pro (Open LLM Leaderboard v2) | 11.6 |
| MuSR (Open LLM Leaderboard v2) | 36.5 |
| OpenBookQA | 38.8 |
| PIQA | 76.7 |
| WinoGrande | 66.1 |