# Evaluation — Qwen3.8-27B · ARCHsloth Q6_K

Every number below is parsed straight out of the log files shipped in [`eval/logs/`](eval/logs). Nothing is retyped.

`llama-perplexity --kl-divergence --kl-divergence-base <bf16 teacher>`, ctx 512, 60 chunks, one H100, `-ngl 99`.

## Korean

| | bytes | Mean KLD | ± | Max KLD | 99% KLD | Median KLD | Same top-p | RMS Δp |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| **ARCHsloth** | 22,430,999,840 | **0.001385** | ±0.000031 | 0.308792 | 0.012412 | 0.000702 | 98.222 ± 0.107 | 1.043 ± 0.022 |
| unsloth `UD-Q6_K` | 21,983,677,344 | 0.002195 | ±0.000059 | 0.620269 | 0.022048 | 0.000904 | 97.693 ± 0.121 | 1.334 ± 0.036 |
| unsloth `UD-Q6_K_M` | 23,088,409,504 | 0.001632 | ±0.000033 | 0.200979 | 0.016601 | 0.000681 | 98.020 ± 0.113 | 1.158 ± 0.032 |

Ours against each, Mean KLD: unsloth `UD-Q6_K` **−36.9 %** · unsloth `UD-Q6_K_M` **−15.1 %**

## English

| | bytes | Mean KLD | ± | Max KLD | 99% KLD | Median KLD | Same top-p | RMS Δp |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| **ARCHsloth** | 22,430,999,840 | **0.001254** | ±0.000026 | 0.111978 | 0.011213 | 0.000638 | 98.183 ± 0.108 | 1.042 ± 0.030 |
| unsloth `UD-Q6_K` | 21,983,677,344 | 0.001476 | ±0.000033 | 0.143187 | 0.015198 | 0.000586 | 98.144 ± 0.109 | 1.106 ± 0.032 |
| unsloth `UD-Q6_K_M` | 23,088,409,504 | 0.001137 | ±0.000029 | 0.158060 | 0.011424 | 0.000454 | 98.314 ± 0.104 | 0.982 ± 0.035 |

Ours against each, Mean KLD: unsloth `UD-Q6_K` **−15.0 %** · unsloth `UD-Q6_K_M` **+10.3 %**