WikiText-2
A commonly used language modeling benchmark dataset derived from Wikipedia text. It is frequently used to evaluate model perplexity.
Timeline (last 90 days)
Decoding every WikiText-2 token through the INT4 weight path raises perplexity from 6.24 to 7.23 (round-to-nearest weights) or 6.99 (GPTQ in the same format).
When doing 50% pruning on Qwen3-4B, CACTUS raised the compressed-dense per-sample gradient cosine similarity on WikiText-2 from 0.670 to 0.755.
In WikiText-2 GGUF runtime testing, Q8_0 quantization retained 99.1% of the BF16 reader NLL gain.
All Facts (3)
Decoding every WikiText-2 token through the INT4 weight path raises perplexity from 6.24 to 7.23 (round-to-nearest weights) or 6.99 (GPTQ in the same format).
50%UnverifiedWhen doing 50% pruning on Qwen3-4B, CACTUS raised the compressed-dense per-sample gradient cosine similarity on WikiText-2 from 0.670 to 0.755.
50%UnverifiedIn WikiText-2 GGUF runtime testing, Q8_0 quantization retained 99.1% of the BF16 reader NLL gain.
50%