modprompt v0.7.2

Modprompt

A collection of prompt templates for language models

• Classic templates formats for different models
• Easily modify and adapt templates on-the-fly
• Few shots and conversation history support

:books: Api doc

Install

npm install modprompt
# or
yarn add modprompt

Usage

Available templates

List the available templates:

import { templates } from "modprompt";

const templateNames = Object.keys(templates);
console.log(templateNames);

Load a template

To load a template:

import { templates, PromptTemplate } from "modprompt";

const tpl = new PromptTemplate(templates.alpaca)
// or
const tpl = new PromptTemplate("alpaca")

Render a template

To render a template to string:

const templateText = tpl.render();

Rendering for an Alpaca template:

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{prompt}

### Response:

Render a prompt

Similar to the render function above, but replaces {prompt} by the provided text:

const templateText = tpl.prompt("What is the capital of Kenya?");

System messages

The template have system messages support if the original template supports it.

To replace a system message:

tpl.replaceSystem("You are a javascript specialist");

Rendering for an Alpaca template:

You are a javascript specialist

### Instruction:
{prompt}

### Response:

To append to a system message:

tpl.afterSystem("You are a javascript specialist");

Note: some templates does have a system schema but no default system message. Some templates don't even have a system block. The default render will show the system schema: exemple for the Vicuna system template:

SYSTEM: {system}

USER: {prompt}

### ASSISTANT:

In case of empty system message it is possible to skip it using the skip_empty_system option: outptut of tpl.render(true):

USER: {prompt}

### ASSISTANT:

Example shots

The templates have support for example shots. Add one shot:

tpl.addShot(
  "fix this invalid json:\n\n```json\n{'a':1,}\n```",
  '\n\n```json\n{"a":1}\n```\n',
)

The first param is the user question, and the second is the assistant response. It is assembled in a prompt template. Example with custom user and assistant messages:

// modify system message
tpl.afterSystem("You are a javascript specialist");
// modify assistant message
tpl.afterAssistant(" (answer in valid json)")
// modify the prompt message
tpl.replacePrompt("fix this invalid json:\n\n```json\n{prompt}\n```")
// add a one shot example
tpl.addShot(
  "{'a':1,}",
  '\n\n```json\n{"a":1}\n```\n',
)

Rendering for an Alpaca template:

Below is an instruction that describes a task. Write a response that appropriately completes the request. You are a javascript specialist

### Instruction:
fix this invalid json:

'''json
{'a':1,}
'''

### Response: (answer in valid json)

'''json
{"a":1}
'''

### Instruction:
fix this invalid json:

'''json
{prompt}
'''

### Response: (answer in valid json)

Chainable api

The calls can be chained. Example with the code above:

const tpl = new PromptTemplate(templates.llama)
  .afterSystem("You are a javascript specialist")
  .afterAssistant(" (answer in valid json)")
  .replacePrompt("fix this invalid json:\n\n```json\n{prompt}\n```")
  .addShot(
    "{'a':1,}",
    '\n\n```json\n{"a":1}\n```\n',
  );

Types

Template types:

interface SpacingSlots {
  system?: number;
  user?: number;
  assistant?: number;
}

interface PromptBlock {
  schema: string;
  message?: string;
}

interface TurnBlock {
  user: string;
  assistant: string;
}

interface LmTemplate {
  name: string;
  user: string;
  assistant: string;
  system?: PromptBlock;
  shots?: Array<TurnBlock>;
  stop?: Array<string>;
  linebreaks?: SpacingSlots;
}

Example raw template:

import type { LmTemplate } from "modprompt";

const orca: LmTemplate = {
  "name": "Orca",
  "system": {
    "schema": "### System:\n{system}",
    "message": "You are an AI assistant that follows instruction extremely well. Help as much as you can.",
  },
  "user": "### User:\n{prompt}",
  "assistant": "### Response:",
  "linebreaks": {
    "system": 2,
    "user": 2
  }
}

:books: Api doc