WEBVTT 1 00:00:00.630 --> 00:00:02.640 So we loaded a model 2 00:00:02.640 --> 00:00:04.350 and we're able to chat with it. 3 00:00:04.350 --> 00:00:06.360 And you can always start new chat sessions 4 00:00:06.360 --> 00:00:08.430 by clicking this plus button here of course. 5 00:00:08.430 --> 00:00:11.130 But let's stick to this one chat session we have here 6 00:00:11.130 --> 00:00:13.650 because you got a couple of things you can configure 7 00:00:13.650 --> 00:00:16.020 regarding this session. 8 00:00:16.020 --> 00:00:18.120 For one, you got some controls down here. 9 00:00:18.120 --> 00:00:21.060 You could, for example, regenerate the message. 10 00:00:21.060 --> 00:00:23.100 So if you're not happy with the response you got, 11 00:00:23.100 --> 00:00:24.810 but you don't wanna change your prompt, 12 00:00:24.810 --> 00:00:27.300 you can click this button to get a new output, 13 00:00:27.300 --> 00:00:30.300 because of course the output will not always be the same 14 00:00:30.300 --> 00:00:32.910 because as explained earlier in the course, 15 00:00:32.910 --> 00:00:35.580 for every token that's generated by the model, 16 00:00:35.580 --> 00:00:38.370 a probability will be assigned to that token, 17 00:00:38.370 --> 00:00:41.040 and then it simply depends on how you configured 18 00:00:41.040 --> 00:00:42.420 for a token to be chosen. 19 00:00:42.420 --> 00:00:45.120 It's not always the one with the highest probability. 20 00:00:45.120 --> 00:00:46.860 So there is some randomness to that, 21 00:00:46.860 --> 00:00:48.990 and that's why you can regenerate a message 22 00:00:48.990 --> 00:00:52.200 to get a different result for the same prompt. 23 00:00:52.200 --> 00:00:54.030 You can also copy the output of course, 24 00:00:54.030 --> 00:00:56.370 and you can edit it. 25 00:00:56.370 --> 00:00:59.340 So you could fine-tune the response you got, 26 00:00:59.340 --> 00:01:00.720 and since it will still be part 27 00:01:00.720 --> 00:01:02.550 of your overall chat history, 28 00:01:02.550 --> 00:01:05.370 future messages you send to that model 29 00:01:05.370 --> 00:01:07.200 would take that edited message 30 00:01:07.200 --> 00:01:08.430 into account as well, 31 00:01:08.430 --> 00:01:11.130 because it's always the entire chat history 32 00:01:11.130 --> 00:01:12.330 of a given chat 33 00:01:12.330 --> 00:01:13.950 that is sent to the model 34 00:01:13.950 --> 00:01:16.623 with every new followup message you send. 35 00:01:17.580 --> 00:01:19.770 You can also delete it if you're not happy with it, 36 00:01:19.770 --> 00:01:23.133 then it would no longer be part of that chat history. 37 00:01:24.030 --> 00:01:27.300 Now, you also got some options up here. 38 00:01:27.300 --> 00:01:29.400 For one, you can duplicate that chat 39 00:01:29.400 --> 00:01:32.310 if you maybe wanna continue with this chat 40 00:01:32.310 --> 00:01:33.480 in one way here, 41 00:01:33.480 --> 00:01:35.310 but send another followup message 42 00:01:35.310 --> 00:01:38.340 for that same chat in another separate chat, 43 00:01:38.340 --> 00:01:40.200 you could then duplicate it. 44 00:01:40.200 --> 00:01:42.960 You could clear it to clear the entire chat history 45 00:01:42.960 --> 00:01:45.420 and get rid of all the messages. 46 00:01:45.420 --> 00:01:47.670 And you can also change the appearance. 47 00:01:47.670 --> 00:01:49.140 So you can configure 48 00:01:49.140 --> 00:01:51.450 whether it should actually be parsed as markdown 49 00:01:51.450 --> 00:01:53.550 or if you wanna see the plain text. 50 00:01:53.550 --> 00:01:55.290 So that would now be the plain text 51 00:01:55.290 --> 00:01:56.820 using markdown formatting, 52 00:01:56.820 --> 00:01:59.883 but now not parsed as markdown by LM Studio. 53 00:02:00.900 --> 00:02:04.620 You could also choose Monospace to have a different font. 54 00:02:04.620 --> 00:02:06.000 And you might have different options 55 00:02:06.000 --> 00:02:08.070 at the point of time where you are viewing this, 56 00:02:08.070 --> 00:02:09.540 but for many scenarios, 57 00:02:09.540 --> 00:02:12.390 markdown is of course the option you want. 58 00:02:12.390 --> 00:02:13.980 You can also tweak the font size 59 00:02:13.980 --> 00:02:15.690 to make it a bit smaller. 60 00:02:15.690 --> 00:02:18.240 The font weight can be changed. 61 00:02:18.240 --> 00:02:20.940 And of course you can also change the message style. 62 00:02:20.940 --> 00:02:24.633 So if you wanna have bubbles or block style. 63 00:02:26.130 --> 00:02:28.140 If you are interacting with a reasoning model, 64 00:02:28.140 --> 00:02:29.580 which this one here isn't, 65 00:02:29.580 --> 00:02:32.760 you could also expand reasoning blocks by default, 66 00:02:32.760 --> 00:02:36.180 change the container width of that chat window, 67 00:02:36.180 --> 00:02:38.850 and automatically scroll new messages to the top. 68 00:02:38.850 --> 00:02:41.160 And again, there might be more options 69 00:02:41.160 --> 00:02:42.690 to fine-tune this appearance 70 00:02:42.690 --> 00:02:44.190 when you're viewing this. 71 00:02:44.190 --> 00:02:47.370 You can also set a system prompt 72 00:02:47.370 --> 00:02:49.353 on a per-chat basis. 73 00:02:50.280 --> 00:02:52.740 Now, to make sure we're all on the same page, 74 00:02:52.740 --> 00:02:57.270 a system prompt is essentially some kind of base prompt 75 00:02:57.270 --> 00:03:01.530 that kind of overrides all other prompts, 76 00:03:01.530 --> 00:03:04.260 all other messages in a chat. 77 00:03:04.260 --> 00:03:06.540 Now, that's not entirely true. 78 00:03:06.540 --> 00:03:09.450 It is part of the overall chat history, 79 00:03:09.450 --> 00:03:14.450 but it is treated in a prioritized way by the AI model, 80 00:03:14.520 --> 00:03:16.350 or at least that's the goal. 81 00:03:16.350 --> 00:03:18.180 So typically a system prompt 82 00:03:18.180 --> 00:03:21.600 is used to set up general instructions 83 00:03:21.600 --> 00:03:26.100 the AI model should follow with its responses. 84 00:03:26.100 --> 00:03:27.960 You can almost think of it 85 00:03:27.960 --> 00:03:31.770 as some prompt that's invisibly injected 86 00:03:31.770 --> 00:03:33.780 in front of every other prompt 87 00:03:33.780 --> 00:03:35.190 you might be sending to the model. 88 00:03:35.190 --> 00:03:36.870 Though that's technically not how it works 89 00:03:36.870 --> 00:03:38.070 because as explained, 90 00:03:38.070 --> 00:03:39.450 it's simply one message 91 00:03:39.450 --> 00:03:41.880 that's part of the overall chat history, 92 00:03:41.880 --> 00:03:44.490 but it is treated with priority. 93 00:03:44.490 --> 00:03:45.660 Now, if I click here, 94 00:03:45.660 --> 00:03:48.390 you'll see at the moment there is no system prompt. 95 00:03:48.390 --> 00:03:52.440 So a blank new chat starts without a system prompt. 96 00:03:52.440 --> 00:03:54.630 But if you know that you wanna have a chat message 97 00:03:54.630 --> 00:03:56.820 where certain rules apply 98 00:03:56.820 --> 00:03:59.880 or where the model should behave in a certain way, 99 00:03:59.880 --> 00:04:03.090 you could of course set such a system prompt. 100 00:04:03.090 --> 00:04:06.483 So here I could say, "Always respond in rhymes. 101 00:04:07.980 --> 00:04:11.640 No matter what the user asks you, 102 00:04:11.640 --> 00:04:13.767 respond in rhymes." 103 00:04:15.300 --> 00:04:17.730 If I now save this system prompt here, 104 00:04:17.730 --> 00:04:19.593 it will apply to this chat. 105 00:04:20.520 --> 00:04:22.050 So if I now send a followup message 106 00:04:22.050 --> 00:04:26.217 like, "Please summarize your last message," 107 00:04:28.800 --> 00:04:31.830 it should go ahead and respond in rhymes 108 00:04:31.830 --> 00:04:35.010 because that's what I set up in this system prompt. 109 00:04:35.010 --> 00:04:37.680 So these are some ways of fine-tuning 110 00:04:37.680 --> 00:04:39.480 the behavior of this model 111 00:04:39.480 --> 00:04:41.700 in this specific chat, 112 00:04:41.700 --> 00:04:43.290 of changing the appearance 113 00:04:43.290 --> 00:04:45.210 and maybe duplicating this chat, 114 00:04:45.210 --> 00:04:48.150 and of controlling these different messages, 115 00:04:48.150 --> 00:04:51.633 deleting them, editing them, or regenerating them.