Icodetools commited on
Commit
d6a5fdb
·
verified ·
1 Parent(s): d389224

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +1 -1
README.md CHANGED
@@ -56,7 +56,7 @@ Named after Vulcan, the Roman god of fire and the forge, the smith who crafts to
56
 
57
  | file | quant | size | notes |
58
  |---|---|---|---|
59
- | `Vulcan-1-Coder-Q4_K_M-league.gguf` | Q4_K_M | 1.56 GB | the exact weights that won the selection league - Edit: Found a bug with quantization. It's byte-for-byte the same, but this one was made from BF16 leading to it being completely different. It changed the behaviour, and explains why the Q8 benched worse than the Q4_K_M |
60
 
61
  ## Results
62
 
 
56
 
57
  | file | quant | size | notes |
58
  |---|---|---|---|
59
+ | `Vulcan-1-Coder-Q4_K_M-league.gguf` | Q4_K_M | 1.56 GB | the exact weights that won the selection league - Edit: Found a bug with quantization. It's byte-for-byte the same, but this one was made from BF16 leading to it being completely different. It changed the behaviour, and explains why the Q8 benched worse than the Q4_K_M. Many other models have this exact bug without the author knowing. It's caused by stacking approximation. A quantisation is an approximation. BF16 is an approximation. Approximating approximations means drift happens. More approximations from using another base means more drift. I will explain it all properly with documentation |
60
 
61
  ## Results
62