Skip to content

Commit 6345532

Browse files
mshumermshumer
committed
Created using Colab
1 parent 1e9a20c commit 6345532

File tree

1 file changed

+389
-0
lines changed

1 file changed

+389
-0
lines changed

‎Claude_3_5_Sonnet_to_gpt_4o_mini_Conversion.ipynb

Lines changed: 389 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,389 @@
1+
{
2+
"cells": [
3+
{
4+
"cell_type": "markdown",
5+
"metadata": {
6+
"id": "view-in-github",
7+
"colab_type": "text"
8+
},
9+
"source": [
10+
"<a href=\"https://colab.research.google.com/github/mshumer/gpt-prompt-engineer/blob/main/Claude_3_5_Sonnet_to_gpt_4o_mini_Conversion.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
11+
]
12+
},
13+
{
14+
"cell_type": "markdown",
15+
"metadata": {
16+
"id": "WljjH8K3s7kG"
17+
},
18+
"source": [
19+
"# Claude 3.5 Sonnet to gpt-4o-mini - part of the `gpt-prompt-engineer` repo\n",
20+
"\n",
21+
"This notebook gives you the ability to go from Claude 3.5 Sonnet to GPT-4o-mini -- reducing costs massively while keeping quality high.\n",
22+
"\n",
23+
"By Matt Shumer (https://twitter.com/mattshumer_)\n",
24+
"\n",
25+
"Github repo: https://github.com/mshumer/gpt-prompt-engineer"
26+
]
27+
},
28+
{
29+
"cell_type": "code",
30+
"execution_count": null,
31+
"metadata": {
32+
"id": "dQmMZdkG_RA5"
33+
},
34+
"outputs": [],
35+
"source": [
36+
"!pip install openai\n",
37+
"\n",
38+
"OPENAI_API_KEY = \"YOUR API KEY HERE\" # enter your OpenAI API key here\n",
39+
"ANTHROPIC_API_KEY = \"YOUR API KEY HERE\" # enter your Anthropic API key here"
40+
]
41+
},
42+
{
43+
"cell_type": "code",
44+
"execution_count": null,
45+
"metadata": {
46+
"id": "wXeqMQpzzosx"
47+
},
48+
"outputs": [],
49+
"source": [
50+
"import re\n",
51+
"import json\n",
52+
"import requests\n",
53+
"from openai import OpenAI\n",
54+
"\n",
55+
"client = OpenAI(api_key=OPENAI_API_KEY)\n",
56+
"\n",
57+
"def generate_candidate_prompts(task, prompt_example, response_example):\n",
58+
" headers = {\n",
59+
" \"x-api-key\": ANTHROPIC_API_KEY,\n",
60+
" \"anthropic-version\": \"2023-06-01\",\n",
61+
" \"content-type\": \"application/json\"\n",
62+
" }\n",
63+
"\n",
64+
" data = {\n",
65+
" \"model\": 'claude-3-5-sonnet-20240620',\n",
66+
" \"max_tokens\": 4000,\n",
67+
" \"temperature\": .5,\n",
68+
" \"system\": \"\"\"<task>Given an example training sample, create seven additional samples for the same task that are even better. Each example should contain a <prompt> and a <response>.</task>\n",
69+
"\n",
70+
"<rules>\n",
71+
"1. Ensure the new examples are diverse and unique from one another.\n",
72+
"2. They should all be perfect. If you make a mistake, this system won't work.\n",
73+
"</rules>\n",
74+
"\n",
75+
"Respond in this format:\n",
76+
"<response_format>\n",
77+
"<example_one>\n",
78+
"<prompt>\n",
79+
"PUT_PROMPT_HERE\n",
80+
"</prompt>\n",
81+
"<response>\n",
82+
"PUT_RESPONSE_HERE\n",
83+
"</response>\n",
84+
"</example_one>\n",
85+
"\n",
86+
"<example_two>\n",
87+
"<prompt>\n",
88+
"PUT_PROMPT_HERE\n",
89+
"</prompt>\n",
90+
"<response>\n",
91+
"PUT_RESPONSE_HERE\n",
92+
"</response>\n",
93+
"</example_two>\n",
94+
"\n",
95+
"...\n",
96+
"</response_format>\"\"\",\n",
97+
" \"messages\": [\n",
98+
" {\"role\": \"user\", \"content\": f\"\"\"<training_task>{task}</training_task>\n",
99+
"\n",
100+
"<prompt_example>\n",
101+
"{prompt_example}\n",
102+
"</prompt_example>\n",
103+
"\n",
104+
"<response_example>\n",
105+
"{response_example}\n",
106+
"</response_example>\"\"\"},\n",
107+
" ]\n",
108+
" }\n",
109+
"\n",
110+
"\n",
111+
" response = requests.post(\"https://api.anthropic.com/v1/messages\", headers=headers, json=data)\n",
112+
"\n",
113+
" response_text = response.json()['content'][0]['text']\n",
114+
"\n",
115+
" # Parse out the prompts and responses\n",
116+
" prompts_and_responses = []\n",
117+
" examples = re.findall(r'<example_\\w+>(.*?)</example_\\w+>', response_text, re.DOTALL)\n",
118+
" for example in examples:\n",
119+
" prompt = re.findall(r'<prompt>(.*?)</prompt>', example, re.DOTALL)[0].strip()\n",
120+
" response = re.findall(r'<response>(.*?)</response>', example, re.DOTALL)[0].strip()\n",
121+
" prompts_and_responses.append({'prompt': prompt, 'response': response})\n",
122+
"\n",
123+
" return prompts_and_responses\n",
124+
"\n",
125+
"def generate_system_prompt(task, prompt_examples):\n",
126+
" headers = {\n",
127+
" \"x-api-key\": ANTHROPIC_API_KEY,\n",
128+
" \"anthropic-version\": \"2023-06-01\",\n",
129+
" \"content-type\": \"application/json\"\n",
130+
" }\n",
131+
"\n",
132+
" data = {\n",
133+
" \"model\": 'claude-3-5-sonnet-20240620',\n",
134+
" \"max_tokens\": 1000,\n",
135+
" \"temperature\": .5,\n",
136+
" \"system\": \"\"\"<your_role>Given a user-description of their <task> a set of prompt / response pairs (it'll be in JSON for easy reading) for the types of outputs we want to generate given inputs, write a fantastic system prompt that describes the task to be done perfectly.</your_role>\n",
137+
"\n",
138+
"<rules>\n",
139+
"1. Do this perfectly.\n",
140+
"2. Respond only with the system prompt, and nothing else. No other text will be allowed.\n",
141+
"</rules>\n",
142+
"\n",
143+
"Respond in this format:\n",
144+
"<system_prompt>\n",
145+
"WRITE_SYSTEM_PROMPT_HERE\n",
146+
"</system_prompt>\"\"\",\n",
147+
" \"messages\": [\n",
148+
" {\"role\": \"user\", \"content\": f\"\"\"<task>{task}</task>\n",
149+
"\n",
150+
"<prompt_response_examples>\n",
151+
"{str(prompt_examples)}\n",
152+
"</prompt_response_examples>\"\"\"},\n",
153+
" ]\n",
154+
" }\n",
155+
"\n",
156+
"\n",
157+
" response = requests.post(\"https://api.anthropic.com/v1/messages\", headers=headers, json=data)\n",
158+
"\n",
159+
" response_text = response.json()['content'][0]['text']\n",
160+
"\n",
161+
" # Parse out the prompt\n",
162+
" system_prompt = response_text.split('<system_prompt>')[1].split('</system_prompt>')[0].strip()\n",
163+
"\n",
164+
" return system_prompt\n",
165+
"\n",
166+
"def test_mini(generated_examples, prompt_example, system_prompt):\n",
167+
" messages = [{\"role\": \"system\", \"content\": system_prompt}]\n",
168+
"\n",
169+
" for example in generated_examples:\n",
170+
" messages.append({\"role\": \"user\", \"content\": example['prompt']})\n",
171+
" messages.append({\"role\": \"assistant\", \"content\": example['response']})\n",
172+
"\n",
173+
" messages.append({\"role\": \"user\", \"content\": prompt_example.strip()})\n",
174+
"\n",
175+
" response = client.chat.completions.create(\n",
176+
" model=\"gpt-4o-mini\",\n",
177+
" messages=messages,\n",
178+
" max_tokens=2000,\n",
179+
" temperature=0.5\n",
180+
" )\n",
181+
"\n",
182+
" response_text = response.choices[0].message.content\n",
183+
"\n",
184+
" return response_text\n",
185+
"\n",
186+
"def run_mini_conversion_process(task, prompt_example, response_example):\n",
187+
" print('Generating the prompts / responses...')\n",
188+
" # Generate candidate prompts\n",
189+
" generated_examples = generate_candidate_prompts(task, prompt_example, response_example)\n",
190+
"\n",
191+
" print('Prompts / responses generated. Now generating system prompt...')\n",
192+
"\n",
193+
" # Generate the system prompt\n",
194+
" system_prompt = generate_system_prompt(task, generated_examples)\n",
195+
"\n",
196+
" print('System prompt generated:', system_prompt)\n",
197+
"\n",
198+
" print('\\n\\nTesting the new prompt on GPT-4o-mini, using your input example...')\n",
199+
" # Test the generated examples and system prompt with the GPT-4o-mini model\n",
200+
" mini_response = test_mini(generated_examples, prompt_example, system_prompt)\n",
201+
"\n",
202+
" print('GPT-4o-mini responded with:')\n",
203+
" print(mini_response)\n",
204+
"\n",
205+
" print('\\n\\n!! CHECK THE FILE DIRECTORY, THE PROMPT IS NOW SAVED THERE !!')\n",
206+
"\n",
207+
" # Create a dictionary with all the relevant information\n",
208+
" result = {\n",
209+
" \"task\": task,\n",
210+
" \"initial_prompt_example\": prompt_example,\n",
211+
" \"initial_response_example\": response_example,\n",
212+
" \"generated_examples\": generated_examples,\n",
213+
" \"system_prompt\": system_prompt,\n",
214+
" \"mini_response\": mini_response\n",
215+
" }\n",
216+
"\n",
217+
" # Save the GPT-4o-mini prompt to a Python file\n",
218+
" with open(\"gpt4o_mini_prompt.py\", \"w\") as file:\n",
219+
" file.write('system_prompt = \"\"\"' + system_prompt + '\"\"\"\\n\\n')\n",
220+
"\n",
221+
" file.write('messages = [\\n')\n",
222+
" for example in generated_examples:\n",
223+
" file.write(' {\"role\": \"user\", \"content\": \"\"\"' + example['prompt'] + '\"\"\"},\\n')\n",
224+
" file.write(' {\"role\": \"assistant\", \"content\": \"\"\"' + example['response'] + '\"\"\"},\\n')\n",
225+
"\n",
226+
" file.write(' {\"role\": \"user\", \"content\": \"\"\"' + prompt_example.strip() + '\"\"\"}\\n')\n",
227+
" file.write(']\\n')\n",
228+
"\n",
229+
" return result"
230+
]
231+
},
232+
{
233+
"cell_type": "markdown",
234+
"source": [
235+
"## Fill in your task, prompt_example, and response_example here. Make sure you keep the quality really high here... this is the most important step!"
236+
],
237+
"metadata": {
238+
"id": "ZujTAzhuBMea"
239+
}
240+
},
241+
{
242+
"cell_type": "code",
243+
"source": [
244+
"task = \"refactoring complex code\"\n",
245+
"\n",
246+
"prompt_example = \"\"\"def calculate_total(prices, tax, discount, shipping_fee, gift_wrap_fee, membership_discount):\n",
247+
"\n",
248+
" total = 0\n",
249+
"\n",
250+
" for i in range(len(prices)):\n",
251+
"\n",
252+
" total += prices[i]\n",
253+
"\n",
254+
" if membership_discount != 0:\n",
255+
"\n",
256+
" total = total - (total * (membership_discount / 100))\n",
257+
"\n",
258+
" if discount != 0:\n",
259+
"\n",
260+
" total = total - (total * (discount / 100))\n",
261+
"\n",
262+
" total = total + (total * (tax / 100))\n",
263+
"\n",
264+
" if total < 50:\n",
265+
"\n",
266+
" total += shipping_fee\n",
267+
"\n",
268+
" else:\n",
269+
"\n",
270+
" total += shipping_fee / 2\n",
271+
"\n",
272+
" if gift_wrap_fee != 0:\n",
273+
"\n",
274+
" total += gift_wrap_fee * len(prices)\n",
275+
"\n",
276+
" if total > 1000:\n",
277+
"\n",
278+
" total -= 50\n",
279+
"\n",
280+
" elif total > 500:\n",
281+
"\n",
282+
" total -= 25\n",
283+
"\n",
284+
" total = round(total, 2)\n",
285+
"\n",
286+
" if total < 0:\n",
287+
"\n",
288+
" total = 0\n",
289+
"\n",
290+
" return total\"\"\"\n",
291+
"\n",
292+
"response_example = \"\"\"def calculate_total(prices, tax_rate, discount_rate, shipping_fee, gift_wrap_fee, membership_discount_rate):\n",
293+
"\n",
294+
" def apply_percentage_discount(amount, percentage):\n",
295+
"\n",
296+
" return amount * (1 - percentage / 100)\n",
297+
"\n",
298+
" def calculate_shipping_fee(total):\n",
299+
"\n",
300+
" return shipping_fee if total < 50 else shipping_fee / 2\n",
301+
"\n",
302+
" def apply_tier_discount(total):\n",
303+
"\n",
304+
" if total > 1000:\n",
305+
"\n",
306+
" return total - 50\n",
307+
"\n",
308+
" elif total > 500:\n",
309+
"\n",
310+
" return total - 25\n",
311+
"\n",
312+
" return total\n",
313+
"\n",
314+
" subtotal = sum(prices)\n",
315+
"\n",
316+
" subtotal = apply_percentage_discount(subtotal, membership_discount_rate)\n",
317+
"\n",
318+
" subtotal = apply_percentage_discount(subtotal, discount_rate)\n",
319+
"\n",
320+
"\n",
321+
"\n",
322+
" total = subtotal * (1 + tax_rate / 100)\n",
323+
"\n",
324+
" total += calculate_shipping_fee(total)\n",
325+
"\n",
326+
" total += gift_wrap_fee * len(prices)\n",
327+
"\n",
328+
"\n",
329+
"\n",
330+
" total = apply_tier_discount(total)\n",
331+
"\n",
332+
" total = max(0, round(total, 2))\n",
333+
"\n",
334+
"\n",
335+
"\n",
336+
" return total\"\"\""
337+
],
338+
"metadata": {
339+
"id": "XSZqqOoQ-5_E"
340+
},
341+
"execution_count": null,
342+
"outputs": []
343+
},
344+
{
345+
"cell_type": "markdown",
346+
"source": [
347+
"### Now, let's run this system and get our new prompt! At the end, you'll see a new file pop up in the directory that contains everything you'll need to reduce your costs while keeping quality high w/ gpt-4o-mini!"
348+
],
349+
"metadata": {
350+
"id": "cMO3cJzWA-O0"
351+
}
352+
},
353+
{
354+
"cell_type": "code",
355+
"source": [
356+
"result = run_mini_conversion_process(task, prompt_example, response_example)"
357+
],
358+
"metadata": {
359+
"id": "O-Bn0rupAJqb"
360+
},
361+
"execution_count": null,
362+
"outputs": []
363+
}
364+
],
365+
"metadata": {
366+
"colab": {
367+
"provenance": [],
368+
"include_colab_link": true
369+
},
370+
"kernelspec": {
371+
"display_name": "Python 3",
372+
"name": "python3"
373+
},
374+
"language_info": {
375+
"codemirror_mode": {
376+
"name": "ipython",
377+
"version": 3
378+
},
379+
"file_extension": ".py",
380+
"mimetype": "text/x-python",
381+
"name": "python",
382+
"nbconvert_exporter": "python",
383+
"pygments_lexer": "ipython3",
384+
"version": "3.8.8"
385+
}
386+
},
387+
"nbformat": 4,
388+
"nbformat_minor": 0
389+
}

0 commit comments

Comments
 (0)