🎯 Key points for quick navigation: 00:00 *📈 Improvement in large language model fine-tuning, especially with new models like Llama 3.2, enhances options for AI developers.* 01:11 *📚 Retrieval Augmented Generation (RAG) is simpler than fine-tuning for integrating private knowledge into models.* 02:49 *⚖️ RAG allows easier updating of knowledge bases, while fine-tuning locks model knowledge based on training data.* 04:40 *💡 Choose RAG for basic knowledge integration, but fine-tune for specialized tasks or behaviors.* 05:22 *🔄 Proper data preparation is crucial for successful fine-tuning, with various data sourcing methods available.* 06:59 *🎤 Assembly AI provides accurate transcription services, enhancing the data preparation process for fine-tuning.* 09:33 *🧬 Synthetic data generation can be utilized to create training datasets for specific tasks efficiently.* 10:01 *⚙️ Close-source vs. open-source models: Understand the trade-offs in control, cost, and ease of use during fine-tuning.* 14:11 *🔑 When selecting a base model, consider cost, speed, and the specificity of your use case for optimal performance.* 16:43 *📝 Fine-tuning methods: Full fine-tuning vs. LoRA, with the latter being more efficient and suitable for many applications.* 18:21 *🚀 Using Unsought can drastically reduce memory usage and increase speed in the fine-tuning process on consumer-grade GPUs.* 20:14 *🎯 Focus on training the lag and target modules for optimal model fine-tuning.* 20:28 *⚖️ Finding the right balance for the learning rate is crucial; too high can cause overfitting, while too low may not yield significant changes.* 20:57 *📊 Data preparation involves converting prompts into model-specific syntax using a feature from unsloth called standardized share GPT.* 22:24 *🚀 The fine-tuning process can be efficiently executed using the FFT trainer from Hugging Face, focusing on assistant outputs.* 23:21 *📈 If results aren't satisfactory, consider providing more training data or switching to a larger model for better reasoning.* 24:03 *💾 You can export either the small adapter file or the larger GGF file for deployment, including access to Hugging Face.* 24:30 *🤝 Join the AI Builder club for deeper insights on fine-tuning, deploying models, and collaborating with experienced AI builders.* Made with HARPA AI
You gotta tell us how we can fine-tune an open source model on a non-public PDF. Everybody on RUclips uses hugging face toy datasets for fine-tuning that is basically useless.
Those who introduce fine-tuning on RUclips they definitely know they are just talking nonsense.Even OpenAI and Claude they don't know how to leverage RLHF to solve the real case.for example even you used RLHF for stock trading you definitely will lose your pants!
I use script to make qna using open ai api all my pdf data into csv format, then convert my csv as jsonl for fine tune then make another version into json to upload it to vector database rag purpose, then fine tune gpt4o mini base model, still not give me good result 😅
There are hundreds of new public datasets released on Hugging Face per week and they all work for finetuning. Also if you want it for your specific usecase, everything requires some sort of custom data - e.g. RAG needs a database.
I may sound dumb af, but how is that possible (8:13)? Is it Python or some JSON extension? If an extension, which one? And if not, in what format should it be done? It's not very convenient to separate CoT or formatted text from each other with a separator. Or is it not necessary to take syntax errors into account at all?
hey jason, i've been following you for a while now, and was wondering if your community is beginner friendly or do we need t have some coding notion? thanks.
just feedback! I think you talk too much especially in the beginning which is hard to listen especially the fact that we didn't get any value yet and we get bombarded with just yap. and with you accent its even harder.
@@10xApe I will say that I've tried to use unsloth myself locally and it didn't work. I tried it so many ways and asked questions on their discord but nothing worked. That was about 4 months ago though so maybe they actually made things easier......maybe.
@@RedSky8 Hey Unsloth founder here - what was the error you experienced? FYI we now allow pip install Unsloth which might help with the process. Currently Unsloth does not work on Windows (unless you use WSL2) or Apple but it's coming soon!
@@danielhanchen is there a way to use 8bit quantized model for inference while using unsloth library, we can load the model in 8 bit by sett load_8bit = true and load_4bit = false, but issue will arise when trying to use the model for inference as it will be there a mismatch in dtype "c10::BFloat16 != signed char"
Nice work Jason. I appreciate the pace. Keep up the good work!
This video is like a hidden gem, Thanks Jason!
Very useful. Thank you Jason.
Thanks for the video was waiting on this one😛
Fantastic job of teaching. Fast paced but great.
🎯 Key points for quick navigation:
00:00 *📈 Improvement in large language model fine-tuning, especially with new models like Llama 3.2, enhances options for AI developers.*
01:11 *📚 Retrieval Augmented Generation (RAG) is simpler than fine-tuning for integrating private knowledge into models.*
02:49 *⚖️ RAG allows easier updating of knowledge bases, while fine-tuning locks model knowledge based on training data.*
04:40 *💡 Choose RAG for basic knowledge integration, but fine-tune for specialized tasks or behaviors.*
05:22 *🔄 Proper data preparation is crucial for successful fine-tuning, with various data sourcing methods available.*
06:59 *🎤 Assembly AI provides accurate transcription services, enhancing the data preparation process for fine-tuning.*
09:33 *🧬 Synthetic data generation can be utilized to create training datasets for specific tasks efficiently.*
10:01 *⚙️ Close-source vs. open-source models: Understand the trade-offs in control, cost, and ease of use during fine-tuning.*
14:11 *🔑 When selecting a base model, consider cost, speed, and the specificity of your use case for optimal performance.*
16:43 *📝 Fine-tuning methods: Full fine-tuning vs. LoRA, with the latter being more efficient and suitable for many applications.*
18:21 *🚀 Using Unsought can drastically reduce memory usage and increase speed in the fine-tuning process on consumer-grade GPUs.*
20:14 *🎯 Focus on training the lag and target modules for optimal model fine-tuning.*
20:28 *⚖️ Finding the right balance for the learning rate is crucial; too high can cause overfitting, while too low may not yield significant changes.*
20:57 *📊 Data preparation involves converting prompts into model-specific syntax using a feature from unsloth called standardized share GPT.*
22:24 *🚀 The fine-tuning process can be efficiently executed using the FFT trainer from Hugging Face, focusing on assistant outputs.*
23:21 *📈 If results aren't satisfactory, consider providing more training data or switching to a larger model for better reasoning.*
24:03 *💾 You can export either the small adapter file or the larger GGF file for deployment, including access to Hugging Face.*
24:30 *🤝 Join the AI Builder club for deeper insights on fine-tuning, deploying models, and collaborating with experienced AI builders.*
Made with HARPA AI
Man. This video fucking crushes it. First one if yours I’ve seen and subscribed 3/4 way through. Keep it up 🔥
Awesome work!!
What a nice tutorial. Thank you, it was very useful. At last I am starting to understand something. Great educational work.
Good stuff dude
Hello Jason, could you kindly add the code so that we can follow along.
well done. Good information
Eggxellent Master.
You gotta tell us how we can fine-tune an open source model on a non-public PDF. Everybody on RUclips uses hugging face toy datasets for fine-tuning that is basically useless.
Those who introduce fine-tuning on RUclips they definitely know they are just talking nonsense.Even OpenAI and Claude they don't know how to leverage RLHF to solve the real case.for example even you used RLHF for stock trading you definitely will lose your pants!
Agree. Creating a dataset with sensible data in the cloud makes everything obsolete. But anyhow good explanation!
I use script to make qna using open ai api all my pdf data into csv format, then convert my csv as jsonl for fine tune then make another version into json to upload it to vector database rag purpose, then fine tune gpt4o mini base model, still not give me good result 😅
There are hundreds of new public datasets released on Hugging Face per week and they all work for finetuning. Also if you want it for your specific usecase, everything requires some sort of custom data - e.g. RAG needs a database.
@@yijin241Finetuning works great, if you know what you are doing.
can ya show how to deploy with llama?
I may sound dumb af, but how is that possible (8:13)?
Is it Python or some JSON extension? If an extension, which one? And if not, in what format should it be done? It's not very convenient to separate CoT or formatted text from each other with a
separator. Or is it not necessary to take syntax errors into account at all?
lol it’s shit like this that makes my shit explode watching these ai dudes
hey jason, i've been following you for a while now, and was wondering if your community is beginner friendly or do we need t have some coding notion? thanks.
I believe unsloth doesnt currently support training or inference on 8bit quantized models although it allows u to load it.
I kind of disagree with book example for LoRa. Full finetuning can also be considered as writing parts of book to tailor to your use case
computer specs?
wow
Nice work jason. I am reviewing your content. Would love to pair program and work together. #devgang
just feedback! I think you talk too much especially in the beginning which is hard to listen especially the fact that we didn't get any value yet and we get bombarded with just yap. and with you accent its even harder.
The problem with these misleading videos is that they provide partial information
What do you mean by that ?
@@10xApe I will say that I've tried to use unsloth myself locally and it didn't work. I tried it so many ways and asked questions on their discord but nothing worked. That was about 4 months ago though so maybe they actually made things easier......maybe.
@@RedSky8 Hey Unsloth founder here - what was the error you experienced? FYI we now allow pip install Unsloth which might help with the process.
Currently Unsloth does not work on Windows (unless you use WSL2) or Apple but it's coming soon!
@@danielhanchen is there a way to use 8bit quantized model for inference while using unsloth library, we can load the model in 8 bit by sett load_8bit = true and load_4bit = false, but issue will arise when trying to use the model for inference as it will be there a mismatch in dtype "c10::BFloat16 != signed char"
Focus more on the content, not revenue. Your videos are sometimes good, and sometimes lacking any substance, just space to fill before a advert.
What are you bringing to the table, stop dissing.
Why are you shilling every video? I can no longer trust the tools you recommend