|
| 1 | +# Transfer Learning |
| 2 | + |
| 3 | + |
| 4 | +## What is transfer learning? |
| 5 | + |
| 6 | + |
| 7 | +## Inductive transfer |
| 8 | + |
| 9 | + |
| 10 | +## Transductive transfer |
| 11 | + |
| 12 | + |
| 13 | +## Inductive vs. transductive transfer |
| 14 | + |
| 15 | + |
| 16 | +## Transfer learning, fine-tuning, and prompt engineering |
| 17 | + |
| 18 | + |
| 19 | +## Fine-tuning with a prompt dataset: a practical example |
| 20 | + |
| 21 | + |
| 22 | +```text |
| 23 | +Q. What is SkillUp Academy? |
| 24 | +A. SkillUp Academy is an online learning platform offering a variety of courses across multiple disciplines, enabling learners to gain knowledge and skills at their own pace from anywhere in the world. |
| 25 | +
|
| 26 | +Q. How do I sign up? |
| 27 | +A. Just click on the "Sign Up" button on our homepage, provide your details, and get started! |
| 28 | +
|
| 29 | +Q. How can I pay for a course? |
| 30 | +A. We accept payments through PayPal, and Stripe. |
| 31 | +
|
| 32 | +Q. I am not satisfied with the course. Can I get a refund? |
| 33 | +A. Yes, we offer a 30-day money-back guarantee. If you're not satisfied within this period, you can request a refund. Please see our refund policy for details. |
| 34 | +
|
| 35 | +Q. Where can I find the terms of service for the website? |
| 36 | +A. Our terms of service can be found at the footer of our website or by clicking here. |
| 37 | +
|
| 38 | +Q. Is my personal information secure? |
| 39 | +A. Absolutely. We prioritize your privacy and have stringent measures in place to protect your data. Read more in our privacy policy here. |
| 40 | +
|
| 41 | +Q. Can I learn at my own pace? |
| 42 | +A. Yes, our courses are designed to allow you to learn at your convenience. Once you enroll, you'll have access to the course materials for a specific period, during which you can learn at your own pace. |
| 43 | +
|
| 44 | +Q. I have questions about the course content. How can I get them answered? |
| 45 | +A. Most courses have a discussion forum where you can ask questions, engage with fellow learners, and sometimes get responses from the course instructors. |
| 46 | +
|
| 47 | +Q. Can I access the courses on mobile? |
| 48 | +A. Yes, our platform is mobile-friendly, and we also have a dedicated app available for both Android and iOS. |
| 49 | +
|
| 50 | +Q. I'm an instructor. How can I offer a course on your platform? |
| 51 | +A. We're always looking for knowledgeable instructors. Just head to our 'Become an Instructor' page for details on how to collaborate with us. |
| 52 | +
|
| 53 | +Q. Do you offer certificates upon course completion? |
| 54 | +A. Yes, once you successfully complete a course and pass any required assessments, you will receive a certificate of completion. |
| 55 | +
|
| 56 | +Q. What should I do if I face technical issues? |
| 57 | +A. Please reach out to our support team via the 'Contact Us' page, and they'll be happy to assist you. |
| 58 | +``` |
| 59 | + |
| 60 | + |
| 61 | +```json |
| 62 | +{"prompt": "lorem ipsum dolor sit amet.END_PROMPT", "completion": " consectetur adipiscing elit.\n"} |
| 63 | +{"prompt": "Vivamus vel lacus quis diam vestibulum scelerisque.END_PROMPT", "completion": " Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas.\n"} |
| 64 | +...etc |
| 65 | +``` |
| 66 | + |
| 67 | + |
| 68 | +```json |
| 69 | +{"prompt": "What is SkillUp Academy->", "completion": " SkillUp Academy is an online learning platform offering a variety of courses across multiple disciplines, enabling learners to gain knowledge and skills at their own pace from anywhere in the world.\n"} |
| 70 | +{"prompt": "How do I sign up->", "completion": " Just click on the \"Sign Up\" button on our homepage, provide your details, and get started!\n"} |
| 71 | +{"prompt": "How can I pay for a course->", "completion": " We accept payments through PayPal, and Stripe.\n"} |
| 72 | +{"prompt": "I am not satisfied with the course. Can I get a refund->", "completion": " Yes, we offer a 30-day money-back guarantee. If you're not satisfied within this period, you can request a refund. Please see our refund policy for details.\n"} |
| 73 | +{"prompt": "Where can I find the terms of service for the website->", "completion": " Our terms of service can be found at the footer of our website or by clicking here.\n"} |
| 74 | +{"prompt": "Is my personal information secure->", "completion": " Absolutely. We prioritize your privacy and have stringent measures in place to protect your data. Read more in our privacy policy here.\n"} |
| 75 | +{"prompt": "Can I learn at my own pace->", "completion": " Yes, our courses are designed to allow you to learn at your convenience. Once you enroll, you'll have access to the course materials for a specific period, during which you can learn at your own pace.\n"} |
| 76 | +{"prompt": "I have questions about the course content. How can I get them answered->", "completion": " Most courses have a discussion forum where you can ask questions, engage with fellow learners, and sometimes get responses from the course instructors.\n"} |
| 77 | +{"prompt": "Can I access the courses on mobile->", "completion": " Yes, our platform is mobile-friendly, and we also have a dedicated app available for both Android and iOS.\n"} |
| 78 | +{"prompt": "I'm an instructor. How can I offer a course on your platform->", "completion": " We're always looking for knowledgeable instructors. Just head to our 'Become an Instructor' page for details on how to collaborate with us.\n"} |
| 79 | +{"prompt": "Do you offer certificates upon course completion->", "completion": " Yes, once you successfully complete a course and pass any required assessments, you will receive a certificate of completion.\n"} |
| 80 | +{"prompt": "What should I do if I face technical issues->", "completion": " Please reach out to our support team via the 'Contact Us' page, and they'll be happy to assist you.\n"} |
| 81 | +``` |
| 82 | + |
| 83 | + |
| 84 | +```bash |
| 85 | +pip install openai[datalib] |
| 86 | +``` |
| 87 | + |
| 88 | + |
| 89 | +```bash |
| 90 | +export OPENAI_API_KEY=<your-api-key> |
| 91 | +``` |
| 92 | + |
| 93 | + |
| 94 | +```bash |
| 95 | +openai tools fine_tunes.prepare_data -f data.jsonl |
| 96 | +``` |
| 97 | + |
| 98 | + |
| 99 | +```bash |
| 100 | +Analyzing... |
| 101 | + |
| 102 | +- Your file contains 12 prompt-completion pairs. In general, we recommend having at least a few hundred examples. We've found that performance tends to linearly increase for every doubling of the number of examples |
| 103 | +- All prompts end with suffix `->` |
| 104 | +- All completions end with suffix `\n` |
| 105 | +
|
| 106 | +No remediations found. |
| 107 | +
|
| 108 | +You can use your file for fine-tuning: |
| 109 | +> openai api fine_tunes.create -t "data.jsonl" |
| 110 | +
|
| 111 | +After you’ve fine-tuned a model, remember that your prompt has to end with the indicator string `->` for the model to start generating completions, rather than continuing with the prompt. Make sure to include `stop=["\n"]` so that the generated texts ends at the expected place. |
| 112 | +Once your model starts training, it'll approximately take 2.61 minutes to train a `curie` model, and less for `ada` and `babbage`. Queue will approximately take half an hour per job ahead of you. |
| 113 | +``` |
| 114 | + |
| 115 | + |
| 116 | +```bash |
| 117 | +openai api fine_tunes.create -t "data.jsonl" -m curie --suffix "SkillUpAcademy" |
| 118 | +``` |
| 119 | + |
| 120 | + |
| 121 | +```bash |
| 122 | +Upload progress: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2.47k/2.47k [00:00<00:00, 8.63Mit/s] |
| 123 | +Uploaded file from data.jsonl: file-RdACszHknKQC49hdZdGxpoEA |
| 124 | +Created fine-tune: ft-WEYpsygNbeHy8YvxLCaiyB4i |
| 125 | +Streaming events until fine-tuning is complete... |
| 126 | + |
| 127 | +(Ctrl-C will interrupt the stream, but not cancel the fine-tune) |
| 128 | +[..] Created fine-tune: ft-WEYpsygNbeHy8YvxLCaiyB4i |
| 129 | +[..] Fine-tune costs $0.01 |
| 130 | +[..] Fine-tune enqueued. Queue number: 0 |
| 131 | + |
| 132 | +...etc |
| 133 | +``` |
| 134 | +
|
| 135 | +
|
| 136 | +```bash |
| 137 | +openai api fine_tunes.follow -i ft-WEYpsygNbeHy8YvxLCaiyB4i |
| 138 | +``` |
| 139 | +
|
| 140 | +
|
| 141 | +```bash |
| 142 | +[..] Created fine-tune: ft-WEYpsygNbeHy8YvxLCaiyB4i |
| 143 | +[..] Fine-tune costs $0.01 |
| 144 | +[..] Fine-tune enqueued. Queue number: 0 |
| 145 | +[..] Fine-tune started |
| 146 | +[..] Completed epoch 1/4 |
| 147 | +[..] Completed epoch 2/4 |
| 148 | +[..] Completed epoch 3/4 |
| 149 | +[..] Completed epoch 4/4 |
| 150 | +[..] Uploaded model: curie:ft-org:skillupacademy-2023-08-18-14-32-18 |
| 151 | +[..] Uploaded result file: file-3fxL6gfapDyHqG7cjXDzFFOf |
| 152 | +[..] Fine-tune succeeded |
| 153 | + |
| 154 | +Job complete! Status: succeeded 🎉 |
| 155 | +Try out your fine-tuned model: |
| 156 | + |
| 157 | +openai api completions.create -m curie:ft-faun:skillupacademy-2023-08-18-14-32-18 -p <YOUR_PROMPT> |
| 158 | +``` |
| 159 | +
|
| 160 | +
|
| 161 | +```python |
| 162 | +import openai |
| 163 | +import os |
| 164 | + |
| 165 | +openai.api_key = os.getenv("OPENAI_API_KEY") |
| 166 | + |
| 167 | +model = "curie:ft-faun:skillupacademy-2023-08-18-14-32-18" |
| 168 | +prompt = input("Ask a question: ") |
| 169 | +instruction = f""" |
| 170 | +You are a helpful and smart customer service representative for SkillUp Academy. |
| 171 | +Answer the following questions based on the training data provided. |
| 172 | +If you don't know the answer, just say 'Sorry, I don't know the answer.' |
| 173 | +
|
| 174 | +AI: Hi there. I am Felix, the chatbot. How can I help you today? |
| 175 | +User: {prompt} |
| 176 | +""" |
| 177 | + |
| 178 | +full_prompt = instruction + " " + prompt + "->" |
| 179 | + |
| 180 | +response = openai.Completion.create( |
| 181 | + model=model, |
| 182 | + prompt=full_prompt, |
| 183 | + max_tokens=100, # adjust based on your need |
| 184 | + stop=["\n"], |
| 185 | + temperature=0 # strikes a balance between deterministic and random |
| 186 | +) |
| 187 | + |
| 188 | +print(response.choices[0].text.strip()) |
| 189 | +``` |
| 190 | +
|
| 191 | +
|
| 192 | +## Why is prompt engineering vital for transfer learning and fine-tuning? |
0 commit comments