path: root/extension/react-app/src/util/modelData.ts

import _ from "lodash";

function updatedObj(old: any, pathToValue: { [key: string]: any }) {
  const newObject = _.cloneDeep(old);
  for (const key in pathToValue) {
    if (typeof pathToValue[key] === "function") {
      _.updateWith(newObject, key, pathToValue[key]);
    } else {
      _.updateWith(newObject, key, (__) => pathToValue[key]);
    }
  }
  return newObject;
}

export enum ModelProviderTag {
  "Requires API Key" = "Requires API Key",
  "Local" = "Local",
  "Free" = "Free",
  "Open-Source" = "Open-Source",
}

export const MODEL_PROVIDER_TAG_COLORS: any = {};
MODEL_PROVIDER_TAG_COLORS[ModelProviderTag["Requires API Key"]] = "#FF0000";
MODEL_PROVIDER_TAG_COLORS[ModelProviderTag["Local"]] = "#00bb00";
MODEL_PROVIDER_TAG_COLORS[ModelProviderTag["Open-Source"]] = "#0033FF";
MODEL_PROVIDER_TAG_COLORS[ModelProviderTag["Free"]] = "#ffff00";

export enum CollectInputType {
  "text" = "text",
  "number" = "number",
  "range" = "range",
}

export interface InputDescriptor {
  inputType: CollectInputType;
  key: string;
  label: string;
  placeholder?: string;
  defaultValue?: string | number;
  min?: number;
  max?: number;
  step?: number;
  options?: string[];
  required?: boolean;
  description?: string;
  [key: string]: any;
}

const contextLengthInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "context_length",
  label: "Context Length",
  defaultValue: 2048,
  required: false,
};
const temperatureInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "temperature",
  label: "Temperature",
  defaultValue: undefined,
  required: false,
  min: 0.0,
  max: 1.0,
  step: 0.01,
};
const topPInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "top_p",
  label: "Top-P",
  defaultValue: undefined,
  required: false,
  min: 0,
  max: 1,
  step: 0.01,
};
const topKInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "top_k",
  label: "Top-K",
  defaultValue: undefined,
  required: false,
  min: 0,
  max: 1,
  step: 0.01,
};
const presencePenaltyInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "presence_penalty",
  label: "Presence Penalty",
  defaultValue: undefined,
  required: false,
  min: 0,
  max: 1,
  step: 0.01,
};
const FrequencyPenaltyInput: InputDescriptor = {
  inputType: CollectInputType.number,
  key: "frequency_penalty",
  label: "Frequency Penalty",
  defaultValue: undefined,
  required: false,
  min: 0,
  max: 1,
  step: 0.01,
};
const completionParamsInputs = [
  contextLengthInput,
  temperatureInput,
  topKInput,
  topPInput,
  presencePenaltyInput,
  FrequencyPenaltyInput,
];

const serverUrlInput = {
  inputType: CollectInputType.text,
  key: "server_url",
  label: "Server URL",
  placeholder: "e.g. http://localhost:8080",
  required: false,
};

export interface ModelInfo {
  title: string;
  class: string;
  description: string;
  longDescription?: string;
  icon?: string;
  tags?: ModelProviderTag[];
  packages: ModelPackage[];
  params?: any;
  collectInputFor?: InputDescriptor[];
}

// A dimension is like parameter count - 7b, 13b, 34b, etc.
// You would set options to the field that should be changed for that option in the params field of ModelPackage
export interface PackageDimension {
  name: string;
  description: string;
  options: { [key: string]: { [key: string]: any } };
}

export interface ModelPackage {
  collectInputFor?: InputDescriptor[];
  description: string;
  title: string;
  refUrl?: string;
  tags?: ModelProviderTag[];
  icon?: string;
  params: {
    model: string;
    template_messages?: string;
    context_length: number;
    stop_tokens?: string[];
    prompt_templates?: any;
    replace?: [string, string][];
    [key: string]: any;
  };
  dimensions?: PackageDimension[];
}

enum ChatTemplates {
  "alpaca" = "template_alpaca_messages",
  "llama2" = "llama2_template_messages",
  "sqlcoder" = "sqlcoder_template_messages",
}

const codeLlamaInstruct: ModelPackage = {
  title: "CodeLlama Instruct",
  description:
    "A model from Meta, fine-tuned for code generation and conversation",
  refUrl: "",
  params: {
    title: "CodeLlama-7b-Instruct",
    model: "codellama:7b-instruct",
    context_length: 2048,
    template_messages: ChatTemplates.llama2,
  },
  icon: "meta.svg",
  dimensions: [
    {
      name: "Parameter Count",
      description: "The number of parameters in the model",
      options: {
        "7b": {
          model: "codellama:7b-instruct",
          title: "CodeLlama-7b-Instruct",
        },
        "13b": {
          model: "codellama:13b-instruct",
          title: "CodeLlama-13b-Instruct",
        },
        "34b": {
          model: "codellama:34b-instruct",
          title: "CodeLlama-34b-Instruct",
        },
      },
    },
  ],
};

const llama2Chat: ModelPackage = {
  title: "Llama2 Chat",
  description: "The latest Llama model from Meta, fine-tuned for chat",
  refUrl: "",
  params: {
    title: "Llama2-7b-Chat",
    model: "llama2:7b-chat",
    context_length: 2048,
    template_messages: ChatTemplates.llama2,
  },
  icon: "meta.svg",
  dimensions: [
    {
      name: "Parameter Count",
      description: "The number of parameters in the model",
      options: {
        "7b": {
          model: "llama2:7b-chat",
          title: "Llama2-7b-Chat",
        },
        "13b": {
          model: "llama2:13b-chat",
          title: "Llama2-13b-Chat",
        },
        "34b": {
          model: "llama2:34b-chat",
          title: "Llama2-34b-Chat",
        },
      },
    },
  ],
};

const wizardCoder: ModelPackage = {
  title: "WizardCoder",
  description:
    "A CodeLlama-based code generation model from WizardLM, focused on Python",
  refUrl: "",
  params: {
    title: "WizardCoder-7b-Python",
    model: "wizardcoder:7b-python",
    context_length: 2048,
    template_messages: ChatTemplates.alpaca,
  },
  icon: "wizardlm.png",
  dimensions: [
    {
      name: "Parameter Count",
      description: "The number of parameters in the model",
      options: {
        "7b": {
          model: "wizardcoder:7b-python",
          title: "WizardCoder-7b-Python",
        },
        "13b": {
          model: "wizardcoder:13b-python",
          title: "WizardCoder-13b-Python",
        },
        "34b": {
          model: "wizardcoder:34b-python",
          title: "WizardCoder-34b-Python",
        },
      },
    },
  ],
};

const phindCodeLlama: ModelPackage = {
  title: "Phind CodeLlama (34b)",
  description: "A finetune of CodeLlama by Phind",
  params: {
    title: "Phind CodeLlama",
    model: "phind-codellama",
    context_length: 2048,
    template_messages: ChatTemplates.llama2,
  },
};

const mistral: ModelPackage = {
  title: "Mistral (7b)",
  description:
    "A 7b parameter base model created by Mistral AI, very competent for code generation and other tasks",
  params: {
    title: "Mistral",
    model: "mistral",
    context_length: 2048,
    template_messages: ChatTemplates.llama2,
  },
  icon: "mistral.png",
};

const sqlCoder: ModelPackage = {
  title: "SQLCoder",
  description:
    "A finetune of StarCoder by Defog.ai, focused specifically on SQL",
  params: {
    title: "SQLCoder",
    model: "sqlcoder",
    context_length: 2048,
    template_messages: ChatTemplates.sqlcoder,
  },
  dimensions: [
    {
      name: "Parameter Count",
      description: "The number of parameters in the model",
      options: {
        "7b": {
          model: "sqlcoder:7b",
          title: "SQLCoder-7b",
        },
        "13b": {
          model: "sqlcoder:15b",
          title: "SQLCoder-15b",
        },
      },
    },
  ],
};

const codeup: ModelPackage = {
  title: "CodeUp (13b)",
  description: "An open-source coding model based on Llama2",
  params: {
    title: "CodeUp",
    model: "codeup",
    context_length: 2048,
    template_messages: ChatTemplates.llama2,
  },
};

const osModels = [
  codeLlamaInstruct,
  llama2Chat,
  wizardCoder,
  phindCodeLlama,
  sqlCoder,
  mistral,
  codeup,
];

export const MODEL_INFO: { [key: string]: ModelInfo } = {
  ollama: {
    title: "Ollama",
    class: "Ollama",
    description:
      "One of the fastest ways to get started with local models on Mac or Linux",
    longDescription:
      'To get started with Ollama, follow these steps:\n1. Download from [ollama.ai](https://ollama.ai/) and open the application\n2. Open a terminal and run `ollama pull <MODEL_NAME>`. Example model names are `codellama:7b-instruct` or `llama2:7b-text`. You can find the full list [here](https://ollama.ai/library).\n3. Make sure that the model name used in step 2 is the same as the one in config.py (e.g. `model="codellama:7b-instruct"`)\n4. Once the model has finished downloading, you can start asking questions through Continue.',
    icon: "ollama.png",
    tags: [ModelProviderTag["Local"], ModelProviderTag["Open-Source"]],
    packages: osModels,
    collectInputFor: [...completionParamsInputs],
  },
  llamacpp: {
    title: "llama.cpp",
    class: "LlamaCpp",
    description: "If you are running the llama.cpp server from source",
    longDescription: `llama.cpp comes with a [built-in server](https://github.com/ggerganov/llama.cpp/tree/master/examples/server#llamacppexampleserver) that can be run from source. To do this:
    
1. Clone the repository with \`git clone https://github.com/ggerganov/llama.cpp\`.
2. \`cd llama.cpp\`
3. Download the model you'd like to use and place it in the \`llama.cpp/models\` directory (the best place to find models is [The Bloke on HuggingFace](https://huggingface.co/TheBloke))
4. Run the llama.cpp server with the command below (replacing with the model you downloaded):

\`\`\`shell
.\\server.exe -c 4096 --host 0.0.0.0 -t 16 --mlock -m models/codellama-7b-instruct.Q8_0.gguf
\`\`\`

After it's up and running, you can start using Continue.`,
    icon: "llamacpp.png",
    tags: [ModelProviderTag.Local, ModelProviderTag["Open-Source"]],
    packages: osModels,
    collectInputFor: [...completionParamsInputs],
  }
};