upgrade to llama-3

README.md

@@ -8,7 +8,7 @@ pinned: false
 app_port: 7860
 ---
 
-Generate Hugging Face Spaces using CodeLlama 34b.
+Generate Hugging Face Spaces using meta-llama/Meta-Llama-3-70B-Instruct
 
 The apps may not always work and usually human work is necessary to finish them.
 See this project as "Hugging Face Space templates on steroids".

package-lock.json

@@ -140,9 +140,9 @@
       "integrity": "sha512-/pyBZWSLD2n0dcHE3hq8s8ZvcETHtEuF+3E7XVt0Ig2nvsVQXdghHVcEkIWjy9A0wKfTn97a/PSDYohKIlnP/w=="
     },
     "node_modules/@types/node": {
-      "version": "20.9.0",
-      "resolved": "https://registry.npmjs.org/@types/node/-/node-20.9.0.tgz",
-      "integrity": "sha512-nekiGu2NDb1BcVofVcEKMIwzlx4NjHlcjhoxxKBNLtz15Y1z7MYf549DFvkHSId02Ax6kGwWntIBPC3l/JZcmw==",
+      "version": "20.10.3",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-20.10.3.tgz",
+      "integrity": "sha512-XJavIpZqiXID5Yxnxv3RUDKTN5b81ddNC3ecsA0SoFXz/QU8OGBwZGMomiq0zw+uuqbL/krztv/DINAQ/EV4gg==",
       "dependencies": {
         "undici-types": "~5.26.4"
       }
@@ -891,9 +891,9 @@
       }
     },
     "node_modules/typescript": {
-      "version": "5.2.2",
-      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.2.2.tgz",
-      "integrity": "sha512-mI4WrpHsbCIcwT9cF4FZvr80QUeKvsUsUvKDoR+X/7XHQH98xYD8YHZg7ANtz2GtZt/CBq2QJ0thkGJMHfqc1w==",
+      "version": "5.3.2",
+      "resolved": "https://registry.npmjs.org/typescript/-/typescript-5.3.2.tgz",
+      "integrity": "sha512-6l+RyNy7oAHDfxC4FzSJcz9vnjTKxrLpDG5M2Vu4SHRVNg6xzqZp6LYSR9zjqQTu8DU/f5xwxUdADOkbrIX2gQ==",
       "peer": true,
       "bin": {
         "tsc": "bin/tsc",

public/index.html

@@ -64,8 +64,8 @@
             <div class="flex flex-col text-stone-700 space-y-1 md:space-y-2">
               <p class="text-stone-700">
                 Model used:
-                <a href="https://huggingface.co/codellama/CodeLlama-34b-Instruct-hf" class="underline" target="_blank">
-                  codellama/CodeLlama-34b-Instruct-hf
+                <a href="https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct" class="underline" target="_blank">
+                  meta-llama/Meta-Llama-3-70B-Instruct (you need to accept the terms)
                 </a>
               </p>
               <p>Powered by 🤗 <a href="https://huggingface.co/inference-api" class="underline" target="_blank">Inference API</a></p>

src/{createLlamaPrompt.mts → createLlamaCoderPrompt.mts}

@@ -1,9 +1,9 @@
 // adapted from https://huggingface.co/TheBloke/Llama-2-13B-chat-GPTQ/discussions/5 
-export function createLlamaPrompt(messages: Array<{ role: string, content: string }>) {
+export function createLlamaCoderPrompt(messages: Array<{ role: string, content: string }>) {
   const B_INST = "[INST]", E_INST = "[/INST]";
   const B_SYS = "<<SYS>>\n", E_SYS = "\n<</SYS>>\n\n";
   const BOS = "<s>", EOS = "</s>";
-  const DEFAULT_SYSTEM_PROMPT = "You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.";
+  const DEFAULT_SYSTEM_PROMPT = "You are an expert coding assistant. Always write code as complete as possible, as the user doesn't have hands. You will receive a generous tip if you write correct code, so take a step back and think rationally about the architecture and data logic of your project workflow.";
 
   if (messages[0].role != "system"){
       messages = [

src/generateFiles.mts

@@ -1,6 +1,6 @@
 import { HfInference } from '@huggingface/inference'
 import { RepoFile } from './types.mts'
-import { createLlamaPrompt } from './createLlamaPrompt.mts'
+import { createLlamaCoderPrompt } from './createLlamaCoderPrompt.mts'
 import { parseTutorial } from './parseTutorial.mts'
 import { getGradioApp } from './getGradioApp.mts'
 import { getStreamlitApp } from './getStreamlitApp.mts'
@@ -28,7 +28,7 @@ export const generateFiles = async (
     ? getReactApp(prompt)
     : getWebApp(prompt)
 
-  const inputs = createLlamaPrompt(instructions) + "\nSure! Here are the source files:\n" + prefix
+  const inputs = createLlamaCoderPrompt(instructions) + "\nSure! Here are the source files:\n" + prefix
 
   let isAbortedOrFailed = false
 
@@ -40,8 +40,11 @@ export const generateFiles = async (
     onProgress(prefix)
 
     for await (const output of hf.textGenerationStream({
-      // model: "tiiuae/falcon-180B-chat",
-      model: "codellama/CodeLlama-34b-Instruct-hf",
+
+      model: "meta-llama/Meta-Llama-3-70B-Instruct",
+      // model: "codellama/CodeLlama-34b-Instruct-hf",
+      // model: "ise-uiuc/Magicoder-CL-7B" // not supported by Hugging Face right now (no stream + max token is 250)
+
       inputs,
       parameters: {
         do_sample: true,
@@ -53,7 +56,10 @@ export const generateFiles = async (
         // for  "tiiuae/falcon-180B-chat":
         //  `inputs` tokens + `max_new_tokens` must be <= 8192
         //  error: `inputs` must have less than 4096 tokens.
-        max_new_tokens: 4096,
+
+        // for Llama-3 it is 8192
+        max_new_tokens: 8192,
+        temperature: 0.8,
         return_full_text: false,
       }
     })) {
@@ -63,10 +69,12 @@ export const generateFiles = async (
       // res.write(output.token.text)
       if (
         tutorial.includes('<|end|>')
-      || tutorial.includes('</s>')
-      || tutorial.includes('[ENDINSTRUCTION]')
-      || tutorial.includes('[/TASK]')
-      || tutorial.includes('<|assistant|>')) {
+        || tutorial.includes('<|eot_id|>')
+        || tutorial.includes('<|start_header_id|>assistant<|end_header_id|>')
+        || tutorial.includes('</s>')
+        || tutorial.includes('[ENDINSTRUCTION]')
+        || tutorial.includes('[/TASK]')
+        || tutorial.includes('<|assistant|>')) {
         tutorial = tutorial.replaceAll("</s>", "").replaceAll("<|end|>", "")
         break
       }

src/getGradioApp.mts

@@ -7,7 +7,7 @@ export function getGradioApp(prompt: string) {
     {
       role: "system",
       content: [
-        `You are a Python developer, expert at crafting Gradio applications to deploy to Hugging Face.`,
+        `You are a Python developer, expert at crafting Gradio applications to deploy to Hugging Face. You must generate valid Python code. Don't forget the requirements.txt files!`,
         `Here is an example of a minimal Gradio application:`,
         gradioDoc
       ].filter(item => item).join("\n")
@@ -16,12 +16,16 @@ export function getGradioApp(prompt: string) {
       role: "user",
       content: `Please write, file by file, the source code for a Gradio project.
 
-You are allowed to use (if necessary) the following Python modules:
-- numpy
+You MUST use the following Python modules:
 - gradio (version 3.39.0)
+
+You are free to use (if necessary) the following Python modules:
+- numpy
 - matplotlib
+- diffusers
+- transformers
 
-Don't forget to write a README.md with the following header:
+Don't forget to write a README.md with the following header, or else you will be FIRED:
 \`\`\`
 ---
 license: apache-2.0