|
17 | 17 | % values reduce it. Setting Temperature=0 removes |
18 | 18 | % randomness from the output altogether. |
19 | 19 | % |
20 | | -% TopP - Top probability mass value for controlling the |
| 20 | +% TopP - Top probability mass value for controlling the |
21 | 21 | % diversity of the output. Default value is 1; |
22 | 22 | % lower values imply that only the more likely |
23 | 23 | % words can appear in any particular place. |
24 | 24 | % This is also known as top-p sampling. |
25 | 25 | % |
26 | | -% TopK - Maximum number of most likely tokens that are |
| 26 | +% TopK - Maximum number of most likely tokens that are |
27 | 27 | % considered for output. Default is Inf, allowing |
28 | 28 | % all tokens. Smaller values reduce diversity in |
29 | 29 | % the output. |
|
34 | 34 | % tail-free sampling. Lower values reduce |
35 | 35 | % diversity, with some authors recommending |
36 | 36 | % values around 0.95. Tail-free sampling is |
37 | | -% slower than using TopP or |
38 | | -% TopK. |
| 37 | +% slower than using TopP or TopK. |
39 | 38 | % |
40 | 39 | % StopSequences - Vector of strings that when encountered, will |
41 | 40 | % stop the generation of tokens. Default |
|
81 | 80 | modelName {mustBeTextScalar} |
82 | 81 | systemPrompt {llms.utils.mustBeTextOrEmpty} = [] |
83 | 82 | nvp.Temperature {llms.utils.mustBeValidTemperature} = 1 |
84 | | - nvp.TopP {llms.utils.mustBeValidTopP} = 1 |
85 | | - nvp.TopK (1,1) {mustBeReal,mustBePositive} = Inf |
| 83 | + nvp.TopP {llms.utils.mustBeValidTopP} = 1 |
| 84 | + nvp.TopK (1,1) {mustBeReal,mustBePositive} = Inf |
86 | 85 | nvp.StopSequences {llms.utils.mustBeValidStop} = {} |
87 | 86 | nvp.ResponseFormat (1,1) string {mustBeMember(nvp.ResponseFormat,["text","json"])} = "text" |
88 | 87 | nvp.TimeOut (1,1) {mustBeReal,mustBePositive} = 120 |
|
0 commit comments