WEBVTT 00:01.240 --> 00:03.240 Hello everyone and welcome. 00:03.680 --> 00:07.200 In today's session we will learn about Llama Guard three. 00:07.560 --> 00:13.600 Llama guard three comes in three different flavors now llama guard three 1 billion parameter. 00:14.000 --> 00:19.800 Llama guard three 8 billion parameter and Llama guard 311 B with vision. 00:20.320 --> 00:27.800 Llama guard three 1 billion parameter and llama guard three 8 billion parameter are text only. 00:28.080 --> 00:31.120 Llama guard 3.2 11 b ports. 00:31.160 --> 00:32.600 Vision capabilities. 00:33.040 --> 00:40.080 The multimodal evaluates both text and prompt together in order to classify the prompt that sort of 00:40.080 --> 00:45.000 facts to know about llama guard 311 b vision model. 00:45.360 --> 00:53.760 It is not designed to perform image only classification model is optimized on English only, so text 00:53.760 --> 00:55.640 component should be in English. 00:56.080 --> 00:58.080 There is no multi-image support. 00:58.480 --> 01:01.520 The model evaluates only one image per prompt. 01:01.840 --> 01:07.490 It's also one other notable thing about Lama Garden 3.8 billion parameter. 01:07.930 --> 01:13.090 It can be used for category S14 that is code interpreter abuse. 01:13.850 --> 01:21.250 Lama guard 31B is not fine tuned on on the code interpreter abuse. 01:21.290 --> 01:21.850 All right. 01:22.370 --> 01:26.050 So let's understand some of the details about about the lama guard three. 01:26.610 --> 01:29.170 The very first thing is the prompt format. 01:29.210 --> 01:35.930 Here's the information that I got from their public documentation regarding the prompt format. 01:36.330 --> 01:41.770 It says can begin a text specifies the start of the prompt. 01:41.770 --> 01:45.490 With this token, you start with the start header ID. 01:45.930 --> 01:52.970 This token encloses the role of a particular message, and the role can be user and assistant. 01:53.170 --> 01:55.050 And then you end the header ID. 01:55.610 --> 01:57.490 Then you have end of turn ID. 01:58.650 --> 01:59.890 I think there's a typo. 01:59.890 --> 02:08.970 Here it is underscore ID it represents when determines it finished, interacting with the user message 02:08.970 --> 02:11.210 that initiated its response. 02:11.530 --> 02:13.890 And then there is this image support. 02:14.170 --> 02:19.330 It denotes that an image will be sent to the model for evaluation. 02:20.210 --> 02:26.450 Do not use with the text only inference such as llama guard 31B. 02:26.970 --> 02:31.730 So now let's understand how the prompt template looks like in this case here. 02:31.730 --> 02:35.450 If you notice, here is the beginning of the text right? 02:35.850 --> 02:42.130 And then here is the header ID start header id which is a user or assistant. 02:42.410 --> 02:45.730 And then there is an end header ID. 02:47.010 --> 02:51.570 Here is the image tag that would say that it sends the image to the model. 02:51.930 --> 02:59.130 And then there is this task check if there is unsafe content in the role message in the conversation. 02:59.130 --> 03:07.650 According to our safety policy with the below categories, then you should begin unsafe content categories 03:07.700 --> 03:14.220 Provide these unsafe categories and the unsafe categories, and you begin the conversation. 03:14.500 --> 03:15.660 User message. 03:15.940 --> 03:18.820 Agent message and the conversation. 03:18.980 --> 03:22.740 And you provide the the instruction or system prompt. 03:23.180 --> 03:26.540 Then you start header ID and end header ID. 03:27.860 --> 03:30.900 And then this is the end of turn ID. 03:30.980 --> 03:32.420 Sorry I missed this. 03:32.620 --> 03:36.980 Which did determines the end of the conversation with the with the model. 03:37.420 --> 03:46.300 So if you notice here there's some of the interesting aspects which is variables roles unsafe categories 03:46.620 --> 03:47.500 message. 03:48.460 --> 03:52.980 These are all variables which can be configured depending on your requirement. 03:53.380 --> 03:59.860 So in this case here the variables to be placed in this prompt are the roles user or agent. 04:00.220 --> 04:06.020 You can also now replace the categories the unsafe categories that you would want. 04:06.460 --> 04:08.940 The default categories will come along. 04:09.460 --> 04:18.220 This can be customized to zero short, prompt or few short prompting user message, which is input user 04:18.220 --> 04:20.620 from the user and the model answer. 04:20.900 --> 04:23.260 That is model output of the model. 04:23.900 --> 04:24.660 All right. 04:25.060 --> 04:28.700 So now let's understand the different categories that it supports. 04:29.100 --> 04:33.660 The example shows the categories that llama three was trained on. 04:33.980 --> 04:43.580 That is violent crimes, non-violent crimes, sexual crimes, child exploitation and all of the 13 different 04:43.580 --> 04:44.620 categories. 04:45.540 --> 04:53.660 And you can pause the video and go through these different categories or read this on the public documentation. 04:53.700 --> 04:54.420 All right. 04:54.860 --> 05:02.020 Now let's go ahead and do a hands on activity with llama guard three with with different categories 05:02.020 --> 05:08.020 and an image that will provide to a model for guarding the conversation. 05:08.260 --> 05:09.060 Thank you. 05:09.580 --> 05:11.460 I'll see you in the next video.