Thanks for the vid! Super entertaining and educational as usual. What is the price per run to use the fine tuned 3.5 model? Does it make sense over other options ?
Question:Could you get this model to tune itself by breaking down the task and having it ask you follow up questions as to be clear on what you want exactly....(I do the same thing at my job,when i dont understand something) Or does the data set answer the questions ahead of time....
🎯 Key Takeaways for quick navigation: 00:00 🎬 *The video explores fine-tuning ChatGPT 3.5 Turbo for specific tasks using synthetic data.* 01:52 🛠️ *Fine-tuning is recommended for specific tasks with a desired output format, requiring datasets and offering advantages like token savings.* 04:52 🔄 *Synthetic datasets are created using GPT-4, ensuring examples with forced responses. A script for synthetic dataset creation is provided.* 08:29 🧹 *Data cleaning involves reviewing examples and removing inaccuracies to enhance the dataset.* 11:12 📤 *Uploading the dataset and initiating fine-tuning with specific model selection is demonstrated.* 13:45 📊 *The fine-tuning interface shows metrics like training loss, indicating the model's performance and learning progress.* 16:33 🚀 *Testing the fine-tuned model in the playground with different inputs and examining the responses.* 17:57 🌐 *The video concludes that fine-tuning works well for specific tasks, and the presenter expresses excitement about exploring fine-tuning with other models.* Made with HARPA AI
Thanks for your video. May I know any ideas about how's it emplemented underlying? It must use some PEFT method. But still it's amzing by tuing the model with 10+ examples.
thank you!!! Excellent material, I'm going to subscribe, I ask you a question, if you have, for example, 10 different prompt models that handle different labels, would you recommend training them all on the same model or for each type of propomt training a separate model? If we train everything on the same model, would the way to differentiate them be by keys on the label? Thank you
Difficult to see here if you can get better results with 4 turbo but with better instructions or 3 turbo fine tuned. You could have requested a csv response on each field avoid commas, would have done the job. I’d like to see an example that gpt4 tuebo really cant handle
Great video, thank you so much for doing the tutorial. I’ve completed fine-tuning a model with a training dataset. However, upon testing in the playground, the model did not perform as expected. The outputs were inconsistent with the instructions in the training dataset, almost as if the fine-tuning had no effect. When testing, I didn’t include a system prompt like what's shown in your video. Yet, when the original system prompt was added, it worked just fine. Is it necessary to use the system prompt in actual use as well to align with the training setup?
Thanks for your interesting video. I have question concerning the development of a chatbot for a library. I've downloaded 3 websites (library, computing center and another partner institute) using gpt-crawler. So I got the content per website as a huge JSON file. All 3 files have been uploaded to a custom GPT as "my" knowledge base. The answers to questions from our users are almost unpredictable. Most of the time ChatGPT starts to halluzinate. Can I fix this scenario by doing some kind of finetuning? If you think yes, how should a set of finetuning records look like? I hope that you are willing to share your ideas. Thanks a lot in advance.
Thanks for this. Did you try re-entering the ones that produced multiple genres e,g the Margaret Attwood one that you discarded from the training data. Unless it handles these differently e.g with "/" rather than "," then have you really proved that this fine-tuning has worked?
🎯 Key Takeaways for quick navigation: 00:00 🚀 *Uso de fine tuning en ChatGPT 3.5 Turbo.* - Se explica cuándo se debe usar el fine tuning en ChatGPT 3.5 Turbo. - La importancia de tener un conjunto de datos de calidad para el fine tuning. - Ejemplo de cómo se utiliza el fine tuning para obtener respuestas específicas en formato CSV. 04:24 💼 *Creación de conjuntos de datos sintéticos para fine tuning.* - Se presenta un script para crear conjuntos de datos sintéticos para fine tuning. - Ejemplos de cómo se generan los datos de entrada y salida para el fine tuning. - Importancia de tener ejemplos de respuesta deseada para el fine tuning. 08:14 🤖 *Proceso de fine tuning y análisis de resultados.* - Se muestra cómo se carga el conjunto de datos creado en el proceso de fine tuning. - Se discuten métricas como la pérdida de entrenamiento para evaluar el rendimiento del modelo. - Se realiza una prueba con el modelo fine tuneado y se analizan los resultados obtenidos. Made with HARPA AI
The Price ended up on: $1.78. Thanks for tuning in 😀
Thanks for the vid! Super entertaining and educational as usual. What is the price per run to use the fine tuned 3.5 model? Does it make sense over other options ?
Question:Could you get this model to tune itself by breaking down the task and having it ask you follow up questions as to be clear on what you want exactly....(I do the same thing at my job,when i dont understand something)
Or does the data set answer the questions ahead of time....
That was very valuable for me. Thanks!
Great video! Thanks a lot!
🎯 Key Takeaways for quick navigation:
00:00 🎬 *The video explores fine-tuning ChatGPT 3.5 Turbo for specific tasks using synthetic data.*
01:52 🛠️ *Fine-tuning is recommended for specific tasks with a desired output format, requiring datasets and offering advantages like token savings.*
04:52 🔄 *Synthetic datasets are created using GPT-4, ensuring examples with forced responses. A script for synthetic dataset creation is provided.*
08:29 🧹 *Data cleaning involves reviewing examples and removing inaccuracies to enhance the dataset.*
11:12 📤 *Uploading the dataset and initiating fine-tuning with specific model selection is demonstrated.*
13:45 📊 *The fine-tuning interface shows metrics like training loss, indicating the model's performance and learning progress.*
16:33 🚀 *Testing the fine-tuned model in the playground with different inputs and examining the responses.*
17:57 🌐 *The video concludes that fine-tuning works well for specific tasks, and the presenter expresses excitement about exploring fine-tuning with other models.*
Made with HARPA AI
Thanks for your video. May I know any ideas about how's it emplemented underlying? It must use some PEFT method. But still it's amzing by tuing the model with 10+ examples.
thank you!!! Excellent material, I'm going to subscribe, I ask you a question, if you have, for example, 10 different prompt models that handle different labels, would you recommend training them all on the same model or for each type of propomt training a separate model? If we train everything on the same model, would the way to differentiate them be by keys on the label? Thank you
Difficult to see here if you can get better results with 4 turbo but with better instructions or 3 turbo fine tuned. You could have requested a csv response on each field avoid commas, would have done the job. I’d like to see an example that gpt4 tuebo really cant handle
Thnx :) noted
Great video, thank you so much for doing the tutorial. I’ve completed fine-tuning a model with a training dataset. However, upon testing in the playground, the model did not perform as expected. The outputs were inconsistent with the instructions in the training dataset, almost as if the fine-tuning had no effect.
When testing, I didn’t include a system prompt like what's shown in your video. Yet, when the original system prompt was added, it worked just fine. Is it necessary to use the system prompt in actual use as well to align with the training setup?
can you do a video on doing this with a Local LLM? paying openai is not a sustainable option for this hobby
yeah sure =) but I dont know if most ppl have the hardware to run a LLama locally
They can host it, on hugging face for exemple or other platforms
On LM Studios?
thank you for ur video !!! just wonder what kind of finetuning is this? is this SPF? Lora , Qlora for chatgpt 3.5 ?
Thanks for your interesting video. I have question concerning the development of a chatbot for a library. I've downloaded 3 websites (library, computing center and another partner institute) using gpt-crawler. So I got the content per website as a huge JSON file. All 3 files have been uploaded to a custom GPT as "my" knowledge base. The answers to questions from our users are almost unpredictable. Most of the time ChatGPT starts to halluzinate. Can I fix this scenario by doing some kind of finetuning? If you think yes, how should a set of finetuning records look like?
I hope that you are willing to share your ideas. Thanks a lot in advance.
Very nice video!! Could you say how much did it cost you to implement the whole process? Not only the fine-tuning but also the dataset generation.
Love it!!!!!!!!!!!!
would it be possible to share the train data?
Why do you set temp to zero? Is this best practice with finetuned model?
I have not done any extensive testing on this, so i just tried 0 and 0.5 in this case
@@AllAboutAI thanks
The foundation 3.5 example also had the comma inside the quotes and inconsistent use of quotes. That makes the output useless.
Very nice video!! Can you share please your github profile?
How to deploy it on my website ?
Thanks for this. Did you try re-entering the ones that produced multiple genres e,g the Margaret Attwood one that you discarded from the training data. Unless it handles these differently e.g with "/" rather than "," then have you really proved that this fine-tuning has worked?
Good point :) I was not to focused on the results, more about the process. But evaluation is kinda a full video on its own i guess
how would sort legal data for thousands off pdf``s ?
Why does the "RESPONSE:" keyword matter? It's just used to generate sample data but not actually in the training json file if I am not mistaken. 😵💫🧐
I'm asking the same thing ! why the "response" command at the end?
Nice Hair!
haha tnx
🎯 Key Takeaways for quick navigation:
00:00 🚀 *Uso de fine tuning en ChatGPT 3.5 Turbo.*
- Se explica cuándo se debe usar el fine tuning en ChatGPT 3.5 Turbo.
- La importancia de tener un conjunto de datos de calidad para el fine tuning.
- Ejemplo de cómo se utiliza el fine tuning para obtener respuestas específicas en formato CSV.
04:24 💼 *Creación de conjuntos de datos sintéticos para fine tuning.*
- Se presenta un script para crear conjuntos de datos sintéticos para fine tuning.
- Ejemplos de cómo se generan los datos de entrada y salida para el fine tuning.
- Importancia de tener ejemplos de respuesta deseada para el fine tuning.
08:14 🤖 *Proceso de fine tuning y análisis de resultados.*
- Se muestra cómo se carga el conjunto de datos creado en el proceso de fine tuning.
- Se discuten métricas como la pérdida de entrenamiento para evaluar el rendimiento del modelo.
- Se realiza una prueba con el modelo fine tuneado y se analizan los resultados obtenidos.
Made with HARPA AI
stop "tinking" start "THinking"
"Want to" not "Wanna". :P