Fine-tuning has become one of the most important skills being asked for in AI engineers today, and it is still one of those skills that not a lot of people actually understand deeply. Many people have made a fine-tuning API call once or twice. But when you ask what's actually happening under the hood, the answers get fuzzy very fast. In this video, I break down the different kinds of fine-tuning methodologies so that when you run your first fine-tuning project, you're not just making API calls. You actually understand the methods deeply from a fundamental level. I cover pre-training vs post-training, open weight vs closed source models, Parameter Efficient Fine-Tuning (LoRA, QLoRA), full fine-tuning, and reinforcement fine-tuning (RLHF, DPO, verifiable rewards). I also share when to use each method and the two biggest challenges: data quality and evaluation. Key insight: QLoRA is the default starting point for most teams fine-tuning open source models today. You can fine-tune a 70 billion parameter model on a single H100 with QLoRA. What questions do you have about fine-tuning? Drop them in the comments. If you want to go deeper on Agentic AI, join our Mastering Agentic AI Bootcamp: https://maven.com/aishwarya-srinivasan/mastering-ai-agents Resources: Best guide: https://cloud.google.com/use-cases/fine-tuning-ai-models Fine-Tuning Libraries- Hugging Face TRL: https://huggingface.co/docs/trl Hugging Face PEFT: https://huggingface.co/docs/peft Axolotl: https://github.com/OpenAccess-AI-Collective/axolotl Unsloth: https://github.com/unslothai/unsloth Inference Providers- Hugging Face Inference: https://huggingface.co/inference-api Nebius Token Factory: https://nebius.ai/ Fine-Tuning APIs (Closed Source)- OpenAI Fine-Tuning: https://platform.openai.com/docs/guides/fine-tuning Google Vertex AI: https://cloud.google.com/vertex-ai Anthropic: https://www.anthropic.com/ Papers- LoRA Paper: https://arxiv.org/abs/2106.09685 QLoRA Paper: https://arxiv.org/abs/2305.14314 DPO Paper: https://arxiv.org/abs/2305.18290 Chapters: 00:00 – Fine-Tuning Is the Most In-Demand Skill 00:32 – Who I Am 00:51 – Pre-Training vs Post-Training 02:52 – Open Weight vs Closed Source 04:14 – PEFT: LoRA and QLoRA 05:30 – Full Fine-Tuning 05:56 – Reinforcement Fine-Tuning 07:43 – When to Use What 09:05 – The Gen Academy 10:20 – Closing
ADVERTISEMENT
what a explaination man! I am fine tuning my first model for the first time.
Why did the Anthropic and OpenAI CEOs claim that AI won't replace humans before their companies' IPOs?
Ever heard of RAG?
I am new in GenAi . What is the fundamental difference between RAG and Fine-tuning
Amazing video and well covered. Simple, crisp and very well covered. Thanks a ton, Aishwarya!
Trust me I have been looking this simple video for long time - But this is extremely incredible video
Btw It's not complete guide. For under the hood you should consider explaining the maths behind it. 👍for your efforts
What are the current machines and its configuration requirements to fine tune a 8B model?
But the cost ita like for USA,UK range, not for india. Engieering cost for bootcamp. 2500 DOllors is 3lkahs.. BUt i know way more knoweldge in this course
why not making a. series into this? where u can train a SLM and show us how to do that
Please share the the Decision Tree. I am a student and part time AI Research Intern at a startup, It would really help me get clear idea to choose my fine tuning approaches. Also, Would love to watch a video on Post Fine - tuning - LLM inference (in local and cloud).
One request aishwarya om recent launch of your course on AI Our organisation refund momey on course but only if your course has a assessmeny with grades being given at end of it Can you please consider adding an assessment Lot of corporate employees can join
Nice presentation
Great presentation. I also watched your video on RAG — very well done. I am intrested in your coutrse, but I do feel the course pricing is on the higher side, even with the 25% coupon, especially for individual learners paying out of pocket in the US. A slightly lower entry price could probably attract a much wider audience.
at this point I am actually learning fine-tuning from hugging face documentation but I don't know more sources I can learn fine tuning better and I am just making basic project to get an idea how they work in multi modal system .
Can you suggest me a Good resource for AI System Design, which covers Modern and Traditional AI System
Good video on fine tuning (which is to say - minimizing predicted outputs vs targeted response simply put) methods, but it is also important to note disadvantages of taking this approach...Operability/maintenance of even a smaller fine tune model (for example: updating a fine tuned model requires training) adds significant year over year cost unless it is truly justified for an enterprise and it must be carefully weighed....I would suggest to look here for a pre-cursor video on this topic..../zYGDpG-pTho?si=F6eWPk60ORct6o7R.... in fact, fine tuning could be approached leveraging prompt, context and also, by giving the model the ability to independently find what it needs in the RAG system using tools.
It is very fast and verbose. One has to know it already to understand your 10 mins talk
💞💞💞
Excellent Explanation Aish!!! As usual... Thanks