MaxDevv/Qwen3.8-27b-Q8-Logits-for-kld

Qwen3.8-27B Q8_0 Reference Logits for KLD

Gold Q8_0 reference logits for Qwen3.8-27B, generated with llama-perplexity --save-all-logits on the hy3 calibration corpus.

• Model: unsloth/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf (29.05 GB)
• Corpus: 20 random samples from canada-quant/hy3-w4a16-mtp-calibration (seed 1337), 512-token windows
• Q8_0 PPL on hy3 corpus: 3.9616 +/- 0.131
• Generated with llama.cpp b10470 llama-perplexity (CPU, 48 threads, kvarn3 KV cache)
• Usage: llama-perplexity -m <quant.gguf> -f hy3_corpus.txt -c 512 --kl-divergence --kl-divergence-base ref_q8.logits

Files:

• ref_q8.logits - per-token logits (float32)
• hy3_corpus.txt - the exact corpus used