metadata
license: apache-2.0
tags:
- not-for-all-audiences
- writing
- roleplay
- gguf
- gguf-imatrix
base_model:
- nakodanei/Blue-Orchid-2x7b
model_type: mixtral
quantized_by: Green-Sky
language:
- en
llama.cpp conversion of https://huggingface.co/nakodanei/Blue-Orchid-2x7b/
except for f16 and q8_0, every quant is using the merge.imatrix
merge.imatrix is a merge of kalomaze-group_10_merged.172chunks.imatrix and wiki.train.400chunks.imatrix, which took ~10min + ~20min to calulate on my machine.
full wiki.train would have taken 10h
for more info on imatrix handling see https://github.com/ggerganov/llama.cpp/pull/5302
ppl (512 wiki.test, 300chunks)
| quant | ppl (lower is better) |
|---|---|
| f16(baseline) | 5.8839 +/- 0.05173 |
| q8_0 | 5.8880 +/- 0.05178 |
| q5_k_m | 5.8912 +/- 0.05177 |
| q5_k_m(without-imat) | 5.8893 +/- 0.05174 |
| q4_k_m | 5.9248 +/- 0.05216 |
| q4_k_m(without-imat) | 5.9492 +/- 0.05249 |
| iq3_xxs | 6.1984 +/- 0.05475 |
| iq3_xxs(only-wiki) | 6.1796 +/- 0.05446 |
| iq3_xxs(only-kal) | 6.1984 +/- 0.05475 |
| iq3_xxs(withou-imat) | 6.4228 +/- 0.05756 |
Interesting observations
despite merge.imatrix being different from kalomaze-group_10_merged.172chunks.imatrix, they produce the exact same quantized iq3_xxs model file. (same hash, checked multiple times)
q5_k_m has a lower perplexity with the imatrix. but that probably is caused by kalomaze-group_10_merged diverging enough from wiki.