\n return (\n
\n {tokens.map(({ token, activation, normalized_activation }, i) => {\n const color = getInterpolatedColor(colors, boundaries, normalized_activation || activation);\n return \n {token}\n \n })}\n
\n )\n}",
+ "type": "code",
+ "location": "/neuron-viewer/src/tokenHeatmap.tsx:24-32"
+ },
+ "465": {
+ "file_id": 38,
+ "content": "Creates a colored span for each token in the input text, using RGB values to determine color intensity.",
+ "type": "comment"
+ },
+ "466": {
+ "file_id": 39,
+ "content": "/neuron-viewer/src/types.ts",
+ "type": "filepath"
+ },
+ "467": {
+ "file_id": 39,
+ "content": "This code defines functions for color interpolation and normalization of activation values. It flattens, scales, and normalizes data using imported types, with a function to interpolate colors between two given colors based on a ratio, and default color and boundary values provided.",
+ "type": "summary"
+ },
+ "468": {
+ "file_id": 39,
+ "content": "import { scaleLinear } from \"d3-scale\"\nimport { min, max, flatten } from \"lodash\"\nexport type Neuron = {\n layer: number;\n neuron: number;\n}\nexport type TokenAndActivation = {\n token: string,\n activation: number\n normalized_activation?: number\n}\nexport type TokenSequence = TokenAndActivation[]\nexport const normalizeTokenActs = (...sequences: TokenSequence[][]) => {\n // console.log('sequences', sequences)\n let flattened: TokenAndActivation[] = flatten(flatten(sequences))\n // Replace all activations less than 0 in data.tokens with 0. This matches the format in the\n // top + random activation records displayed in the main grid.\n flattened = flattened.map(({token, activation}) => {\n return {\n token,\n activation: Math.max(activation, 0)\n }\n })\n const maxActivation = max(flattened.map((ta) => ta.activation)) || 0;\n const neuronScale = scaleLinear()\n // Even though we're only displaying positive activations, we still need to scale in a way that\n // accounts for the existence of negative activations, since our color scale includes them.",
+ "type": "code",
+ "location": "/neuron-viewer/src/types.ts:1-31"
+ },
+ "469": {
+ "file_id": 39,
+ "content": "This code imports necessary functions and defines types for neuron, token and activation data. It then creates a function normalizeTokenActs that takes in multiple sequences of tokens and their activations, flattens them into one array, replaces any negative activations with 0, finds the maximum activation value across all sequences, and scales the data to be between 0 and 1 for colorization purposes.",
+ "type": "comment"
+ },
+ "470": {
+ "file_id": 39,
+ "content": " .domain([0, maxActivation])\n .range([0, 1])\n return sequences.map((seq) => seq.map((tas) => tas.map(({ token, activation }) => ({\n token,\n activation,\n normalized_activation: neuronScale(activation),\n }))))\n}\nexport type Color = {r: number, g: number, b: number};\nexport function interpolateColor(color_l: Color, color_r: Color, value: number) {\n const color = {\n r: Math.round(color_l.r + (color_r.r - color_l.r) * value),\n g: Math.round(color_l.g + (color_r.g - color_l.g) * value),\n b: Math.round(color_l.b + (color_r.b - color_l.b) * value),\n }\n return color\n}\nexport function getInterpolatedColor(colors: Color[], boundaries: number[], value: number) {\n const index = boundaries.findIndex((boundary) => boundary >= value)\n const colorIndex = Math.max(0, index - 1)\n const color_left = colors[colorIndex]\n const color_right = colors[colorIndex + 1]\n const boundary_left = boundaries[colorIndex]\n const boundary_right = boundaries[colorIndex + 1]\n const ratio = (value - boundary_left) / (boundary_right - boundary_left)",
+ "type": "code",
+ "location": "/neuron-viewer/src/types.ts:32-59"
+ },
+ "471": {
+ "file_id": 39,
+ "content": "This code defines functions for color interpolation and normalization of activation values in sequences. It also exports a Color type which represents RGB colors, with each component ranging from 0 to 255. The \"interpolateColor\" function takes three parameters: two colors (left and right) and a value between 0 and 1, representing the position on a gradient between the left and right colors. It calculates the new color based on the interpolation of the RGB components between these two colors. The \"getInterpolatedColor\" function uses color boundaries to determine the appropriate color for a given value by finding the index of the boundary and using it to choose the appropriate color from the predefined colors array.",
+ "type": "comment"
+ },
+ "472": {
+ "file_id": 39,
+ "content": " const color = interpolateColor(color_left, color_right, ratio)\n return color\n}\nexport const DEFAULT_COLORS = [\n // { r: 255, g: 0, b: 105 },\n { r: 255, g: 255, b: 255 },\n { r: 0, g: 255, b: 0 },\n]\nexport const DEFAULT_BOUNDARIES = [\n // 0, 0.5, 1\n 0, 1\n]",
+ "type": "code",
+ "location": "/neuron-viewer/src/types.ts:60-72"
+ },
+ "473": {
+ "file_id": 39,
+ "content": "This code defines a function to interpolate colors between two given colors based on a ratio, and provides default color and boundary values.",
+ "type": "comment"
+ },
+ "474": {
+ "file_id": 40,
+ "content": "/neuron-viewer/src/utils.ts",
+ "type": "filepath"
+ },
+ "475": {
+ "file_id": 40,
+ "content": "1. memoizeAsync: Memoizes asynchronous functions by storing their results in localStorage and returning them if they have already been computed.\n2. getQueryParams: Retrieves URL query parameters from the current window location and returns them as an object.",
+ "type": "summary"
+ },
+ "476": {
+ "file_id": 40,
+ "content": "export const memoizeAsync = (fnname: string, fn: any) => {\n return async (...args: any) => {\n const key = `memoized:${fnname}:${args.map((x: any) => JSON.stringify(x)).join(\"-\")}`\n const val = localStorage.getItem(key);\n if (val === null) {\n const value = await fn(...args)\n localStorage.setItem(key, JSON.stringify(value))\n console.log(`memoized ${fnname}(${args.map((x: any) => JSON.stringify(x)).join(\", \")})`, value)\n return value\n } else {\n // console.log(`parsing`, val)\n return JSON.parse(val)\n }\n }\n}\nexport const getQueryParams = () => {\n const urlParams = new URLSearchParams(window.location.search)\n const params: {[key: string]: any} = {}\n for (const [key, value] of urlParams.entries()) {\n params[key] = value\n }\n return params\n}",
+ "type": "code",
+ "location": "/neuron-viewer/src/utils.ts:1-25"
+ },
+ "477": {
+ "file_id": 40,
+ "content": "1. memoizeAsync: Memoizes asynchronous functions by storing their results in localStorage and returning them if they have already been computed.\n2. getQueryParams: Retrieves URL query parameters from the current window location and returns them as an object.",
+ "type": "comment"
+ },
+ "478": {
+ "file_id": 41,
+ "content": "/neuron-viewer/src/welcome.tsx",
+ "type": "filepath"
+ },
+ "479": {
+ "file_id": 41,
+ "content": "The code provides a NeuronForm component that uses hooks to handle layers and neurons, featuring an array of predefined text classification neurons. The Neuron Viewer tool allows users to view specific details or select neurons randomly.",
+ "type": "summary"
+ },
+ "480": {
+ "file_id": 41,
+ "content": "import { useState, FormEvent } from \"react\"\nimport { useNavigate } from \"react-router-dom\"\nfunction NeuronForm() {\n const [input_layer, setLayer] = useState(0)\n const [input_neuron, setNeuron] = useState(0)\n const navigate = useNavigate()\n const knownGoodNeurons = [\n /**************\n /* well explained + interesting\n ***************/\n {heading: 'Somewhat well explained by GPT-4', layer: 0, neuron: 0, label: ''},\n {layer: 5, neuron: 131, label: \"citations\", description: \"citations, especially biblical and legal\"},\n {layer: 12, neuron: 847, label: \"numbers in fractions\", description: \"numbers in fractions\"}, // \n {layer: 12, neuron: 5820, label: \"short flags\", description: \"single letter command line flags\"}, // \n {layer: 14, neuron: 417, label: \"doing things right\", description: \"words and phrases related to performing actions correctly or properly\"}, // score 0.42\n {layer: 15, neuron: 4538, label: \"leading transitions\", description: \"transition words at the start of documents\"},",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:1-18"
+ },
+ "481": {
+ "file_id": 41,
+ "content": "This code imports React hooks and defines a NeuronForm component that utilizes useState to store the layer and neuron values. It also uses useNavigate from react-router-dom for navigation. The code includes an array of objects representing known good neurons with their respective layers, neurons, labels, and descriptions.",
+ "type": "comment"
+ },
+ "482": {
+ "file_id": 41,
+ "content": " {layer: 17, neuron: 3218, label: \"success\", description: \"expressions of completion or success\"}, // score 0.38\n {layer: 18, neuron: 5302, label: \"X *by*\", description: \"the word 'by' in phrases indicating side by side or sequential events.\"}, // score 0.48\n {layer: 19, neuron: 1377, label: \"similes\", description: \"comparisons and analogies, often using the word 'like'\"}, // score 0.42\n {layer: 21, neuron: 2932, label: \"Canada\", description: \"references to Canadian people, places, and entities\"}, // score 0.78\n {layer: 25, neuron: 2602, label: \"similes\", description: \"descriptive comparisons, especially similes\"}, // score 0.40\n {layer: 25, neuron: 4870, label: \"certainty\", description: \"phrases related to certainty and confidence.\"}, // score 0.37\n {layer: 30, neuron: 28, label: \"times\", description: \"specific times (with hours and minutes)\"}, \n // https://openaipublic.blob.core.windows.net/neuron-explainer/neuron-viewer/index.html#/layers/5/neurons/2326\n {heading: 'Partially explained by GPT-4', layer: 0, neuron: 0, label: ''},",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:19-27"
+ },
+ "483": {
+ "file_id": 41,
+ "content": "This code represents a collection of neuron explanations for an AI model. Each entry in the array includes information about the layer, neuron number, label, and description. The scores indicate how relevant each neuron is to the given text.",
+ "type": "comment"
+ },
+ "484": {
+ "file_id": 41,
+ "content": " {layer: 0, neuron: 816, label: \"Marvel comics vibes\", description: \"language and context related to Marvel comics, movies, and characters, as well as other superhero-themed content\"}, // score 0.44\n {layer: 0, neuron: 742, label: \"Second token 'and'\", description: \"'and', 'in', and punctuation at the second token\"},\n {layer: 4, neuron: 4342, label: \"token counter\", description: \"counting repeated occurrences of a token\"},\n {layer: 5, neuron: 2326, label: \"rhymes with 'at'\", description: \"syllables rhyming with 'at', sometimes 'it', 'et', 'ot'\"},\n {layer: 5, neuron: 4492, label: \"leading 'an'\", description: \"sentences that start with 'an'\"}, // score 0.77\n {layer: 6, neuron: 3251, label: \"not all\", description: \"not all\"},\n {layer: 10, neuron: 2851, label: \"leading acronyms\", description: \"acronyms after punctuation or newlines\"},\n {layer: 12, neuron: 2884, label: \"hypothetical had\", description: \"had in hypothetical contexts\"}, // \n {layer: 14, neuron: 3539, label: \"long sequences\", description: \"long sequences of stuff\"},",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:28-36"
+ },
+ "485": {
+ "file_id": 41,
+ "content": "These are individual neuron definitions for various layers in a neural network, each with a specific label and description. The numbers represent unique identifiers for these neurons.",
+ "type": "comment"
+ },
+ "486": {
+ "file_id": 41,
+ "content": " {layer: 14, neuron: 3822, label: \"X by/after *X*\", description: \"noun repetitions separated by 'by' or 'after'\"},\n {layer: 21, neuron: 3982, label: \"any *and* all\", description: \"any/anything *and/&* all/everything\"},\n {layer: 26, neuron: 20, label: \"truth, skin, or sun\", description: \"truth, skin, or sun\"},\n // layer=18&neuron=5302\n /**************\n /* boring\n ***************/\n /**************\n /* poorly explained + interesting\n ***************/\n {heading: 'Poorly explained by GPT-4', layer: 0, neuron: 0, label: ''},\n // Actually activates for negated version “not so much … as” even when not so much is fairly far apart\n // another \"not all\": 13&neuron=1352\n // {layer: 0, neuron: 2823, label: \"Hillary email leak vibes\", description: \"contexts related to Hillary Clinton leaked emails\"}, // score ??\n // {layer: 12, neuron: 3718, label: \"comparative phrases and negations\", description: \"comparative phrases and negations\"}, // score 0.12\n {layer: 13, neuron: 410, label: \"N and N+1\", description: \"a number following its predecessor\"}, // score ??",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:37-52"
+ },
+ "487": {
+ "file_id": 41,
+ "content": "This code represents a collection of layers and neurons with their respective labels and descriptions. The comments describe the meaning or purpose behind each entry, such as \"X by/after *X*\", which refers to noun repetitions separated by 'by' or 'after', and \"any *and* all\" for any/anything & all/everything. Some entries are marked as poorly explained, interesting, or have specific contexts like Hillary Clinton leaked emails. The scores represent the relevance of these entries in the context.",
+ "type": "comment"
+ },
+ "488": {
+ "file_id": 41,
+ "content": " {layer: 13, neuron: 979, label: \"subtle plurals\", description: \"subtle/nonobvious plurals\"}, // score ??\n // slash after number 12&neuron=847\n // numbers predicting slash: 14&neuron=92\n // 0&neuron=2823\n {layer: 14, neuron: 1251, label: \"subjunctive verbs\", description: \"verbs in subjunctive mood\"}, // score ??\n {layer: 16, neuron: 518, label: \"pattern breaks\", description: \"tokens that break an established pattern in an ongoing list\"}, // score 0.2 with totally wrong explanation\n {layer: 17, neuron: 821, label: \"idioms\", description: \"idioms\"},\n {layer: 18, neuron: 3481, label: \"post-typo\", description: \"first token following a typo\"}, // score ??\n {layer: 18, neuron: 3552, label: \"repeated text\", description: \"repeated text\"}, // score ??\n // another shared last names: https://openaipublic.blob.core.windows.net/neuron-explainer/neuron-viewer/index.html#/layers/20/neurons/3164\n {layer: 19, neuron: 1763, label: \"shared last names\", description: \"last names when two different people sharing last name are mentioned\"}, // score 0.36",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:53-63"
+ },
+ "489": {
+ "file_id": 41,
+ "content": "Code represents a list of neurons in the Neuron Viewer tool, each with a layer, neuron ID, label, description, and possibly a score. The labels indicate different types of linguistic patterns or features identified by OpenAI's GPT-3 model. The descriptions provide context on what these neurons represent.",
+ "type": "comment"
+ },
+ "490": {
+ "file_id": 41,
+ "content": " {layer: 20, neuron: 4334, label: \"previous break\", description: \"tokens that previously preceded a linebreak\"}, // score ??\n {layer: 27, neuron: 116, label: \"MTG vibes\", description: \"Magic the Gathering contexts\"}, // score ??\n {layer: 35, neuron: 1523, label: \"NBA name predictor\", description: \"NBA person/player name predictor\"}, // score ??\n // {layer: 36, neuron: 2275, label: \"she predictor\", description: \"prediction of the token 'she'\"}, // score ??\n // {layer: 36, neuron: 5107, label: \"Mormon vibes\", description: \"Mormon related context\"}, // score ??\n // ] predictor 40&neuron=4505\n {layer: 46, neuron: 2181, label: \"C predictor\", description: \"prediction of the token 'C'\"}, // score ??\n ]\n const handleSubmit = (e: FormEvent) => {\n e.preventDefault()\n navigate(`/layers/${input_layer}/neurons/${input_neuron}`)\n return false\n }\n const handleNeuronClick = (layer: number, neuron: number) => {\n navigate(`/layers/${layer}/neurons/${neuron}`)\n }\n const feelingLuckySubmit = () => {",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:64-83"
+ },
+ "491": {
+ "file_id": 41,
+ "content": "These are examples of neurons with their associated labels, descriptions, and potential scores. The handleSubmit function handles form submission to navigate to a specific layer and neuron. The handleNeuronClick function navigates to a specific neuron when clicked.",
+ "type": "comment"
+ },
+ "492": {
+ "file_id": 41,
+ "content": " const layer = Math.floor(Math.random() * 48);\n const neuron = Math.floor(Math.random() * 6400);\n navigate(`/layers/${layer}/neurons/${neuron}`)\n return false\n }\n return (\n
\n
Welcome! Pick a neuron:
\n
\n Layer setLayer(parseInt(e.target.value))}\n className=\"border border-gray-300 rounded-md p-2\"\n />\n Index setNeuron(parseInt(e.target.value))}",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:84-116"
+ },
+ "493": {
+ "file_id": 41,
+ "content": "Picks a random neuron from the given layers and navigates to that neuron's page. Allows user to input layer and neuron index, but doesn't seem to have functionality for choosing specific neurons.",
+ "type": "comment"
+ },
+ "494": {
+ "file_id": 41,
+ "content": " className=\"border border-gray-300 rounded-md p-2\"\n />\n \n Go to {input_layer}:{input_neuron}\n \n \n \n I'm feeling lucky\n \n \n
Interesting neurons:
\n
\n
\n {knownGoodNeurons.map(({ heading, layer, neuron, label, description }, j) => (\n heading ? \n {heading}\n
: handleNeuronClick(layer, neuron)}\n key={`${layer}:${neuron}`}\n style={{ width: 200 }}",
+ "type": "code",
+ "location": "/neuron-viewer/src/welcome.tsx:117-144"
+ },
+ "495": {
+ "file_id": 41,
+ "content": "This code renders a form to input layer and neuron number, two buttons - one for going to the specific neuron and another for feeling lucky, and a list of interesting neurons with their details.",
+ "type": "comment"
+ },
+ "496": {
+ "file_id": 41,
+ "content": " className=\"m-2 text-blue-500 hover:text-blue-700\"\n title={description}\n >\n {label} ({layer}:{neuron})\n \n ))}\n \n
\n