WikiText-2
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
WikiText-2 Top 5% most uncertain tokens (val)
5.13NLL
5
WikiText-2 All tokens (test)
1.66NLL
5
WikiText-2 All tokens (val)
1.7NLL
5
WikiText-2 1M
1.275Loss
4
WikiText-2 GPT-2-small
37.05Perplexity
4
WikiText-2 (val)
5.083BPB
4
WikiText-2 (held-out query)
0.979Spearman (LDS)
3
WikiText-2 byte-level (val)
3.595BPB
3
WikiText-2 raw v1 (train)
3.574Loss
3
WikiText-2 LLaMA-2 7B
5.12PPL
3
WikiText-2 (val)
3.49PPL (Validation)
3
WikiText-2 1024 sequence length
4.502Perplexity (PPL)
2
WikiText-2 512 sequence length
4.339Perplexity (PPL)
2
WikiText-2 byte
3.918Perplexity
2
WikiText-2 char
3.925Perplexity
2
WikiText-2 32768 context length (val)
2.048Loss
2