DavidAU commited on
Commit
9b969cf
·
verified ·
1 Parent(s): d7b6b9b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +5 -4
README.md CHANGED
@@ -46,10 +46,7 @@ pipeline_tag: text-generation
46
 
47
  It is the new "Mistral-Small-Instruct 2409 22B", max context of 131,000 (128k) with the NEO IMATRIX dataset.
48
 
49
- This model IS bullet proof and operates with all parameters, including temp settings from 0 to 5.
50
-
51
- It is an extraordinary compressed model at a PPL level of 4.8611 +/- 0.06701 (Q4_K_M).
52
-
53
  The NEO IMATRIX dataset V2 was applied to it to enhance creativity.
54
 
55
  4 examples provided to show differences at TEMP=0 and at TEMP=1 for both non-imatrix and NEO imatrix versions.
@@ -64,6 +61,10 @@ Please refer to the original model card for this model from MistralAI for additi
64
 
65
  Imatrix quants perform best at IQ3s and IQ4s, then Q4s, lower on Q5, and tappers off at Q6.
66
 
 
 
 
 
67
  Due to the parameter count of this model, even IQ2s quants will work very well.
68
 
69
  Q8 is not uploaded here because Imatrix has no effect on this quant.
 
46
 
47
  It is the new "Mistral-Small-Instruct 2409 22B", max context of 131,000 (128k) with the NEO IMATRIX dataset.
48
 
49
+ This model IS bullet proof and operates with all parameters, including temp settings from 0 to 5. It is an extraordinary compressed model at a PPL level of 4.8611 +/- 0.06701 (Q4_K_M).
 
 
 
50
  The NEO IMATRIX dataset V2 was applied to it to enhance creativity.
51
 
52
  4 examples provided to show differences at TEMP=0 and at TEMP=1 for both non-imatrix and NEO imatrix versions.
 
61
 
62
  Imatrix quants perform best at IQ3s and IQ4s, then Q4s, lower on Q5, and tappers off at Q6.
63
 
64
+ Recommend: IQ4_XS for maximum imatrix effect and best "bit count".
65
+
66
+ For stronger IMATRIX effect, IQ3s, and IQ2s.
67
+
68
  Due to the parameter count of this model, even IQ2s quants will work very well.
69
 
70
  Q8 is not uploaded here because Imatrix has no effect on this quant.