How to Train Your Own Large Language Models

Given the success of OpenAI’s GPT-4 and Google’s PaLM, every company is now assessing its own use cases for Large Language Models (LLMs). Many companies will ultimately decide to train their own LLMs for a variety of reasons, ranging from data privacy to increased control over updates and improvements. One of the most common reasons will be to make use of proprietary internal data.
In this session, we’ll go over how to train your own LLMs, from raw data to deployment in a user-facing production environment. We’ll discuss the engineering challenges, and the vendors that make up the modern LLM stack: Databricks, Hugging Face, and MosaicML. We’ll also break down what it means to train an LLM using your own data, including the various approaches and their associated tradeoffs.
Topics covered in this session:
- How Replit trained a state-of-the-art LLM from scratch
- The different approaches to using LLMs with your internal data
- The differences between fine-tuning, instruction tuning, and RLHF
Talk by: Reza Shabani
Here’s more to explore:
LLM Compact Guide: dbricks.co/43WuQyb
Big Book of MLOps: dbricks.co/3r0Pqiz
Connect with us: Website: databricks.com
Twitter: / databricks
LinkedIn: / databricks
Instagram: / databricksinc
Facebook: / databricksinc

[@BlackThorne]

00:35 🧠 Business Use Case: Training Large Language Models (LLMs)
01:45 💡 Reasons for Training LLMs: Customization, Cost Efficiency
02:40 🔀 Training Process: Data Pipelines, Model Training, Inference
05:16 📊 Data Processing: Filtering, Anonymization, Pre-processing
08:23 🔤 Tokenizer & Vocabulary Training: Custom Vocabulary, Benefits, Challenges
13:09 🎯 Model Evaluation: Human Eval Framework, Code Metrics vs. NLP Metrics
18:35 ⚙ Model Training & Specs: Model Size, Training Objective, Attention Mechanisms
20:55 📈 Model Training Challenges: Data Determinism, Loss Curve Spikes
23:41 🔄 Generation vs. Evaluation: Separating the Process
24:08 🚀 Deployment: Building Inference Stack, Managed Services
24:52 🖥 Model training involves GPU and model size considerations, pre/post-processing, and server/client-side logic.
25:49 🧠 Evaluating your model is crucial; define success criteria early to guide the training process.
26:02 🔄 Rapid iteration is valuable for testing model behavior and improving user experience.
26:29 ⏳ Ensure compatibility between training and inference stacks to avoid sub-optimal results.
26:57 🔄 Customization drives the desire to train LLMs with one's data; various approaches exist.
27:51 📚 Retrieval-based augmentation involves fetching relevant context to guide model responses.
28:08 🤖 Contextual prompting improves model's domain-specific knowledge, even if not originally trained.
28:21 💡 Embeddings and semantic similarity prioritize context selection for retrieval.
30:01 🎯 Fine-tuning methods vary in complexity; consider instruction tuning and training from scratch.
31:25 🔄 Models struggle with varying data formats, short-form content, and changing facts.
32:35 🌍 Custom domain data presents challenges; careful selection and use of embeddings is key.
33:37 🌶 Fine-tuning is complex; unsupervised fine-tuning for new domain knowledge has limitations.
35:18 🚫 Agents might become redundant as models absorb useful functionalities.
36:00 🔄 Balancing training data mix is challenging; no established formula, lots of variables.
37:11 💾 Data iteration tools are crucial as data, not GPUs, becomes the bottleneck for model advancement.

[@syednaveed1391]

Super useful. I am a physician, tried fine tune using cancer documents. It didn't work. Found your video Thanks

[@dilipjha08]

Thanks for knowledge sharing to the technology user. It was very details about the dlt as well as streaming tables and comprison between it and demo of the topic was very perfect.

[@CalebFenton]

Thanks for the info and esp the hot takes.

[@AnandShah-ds]

That was the best part. Should have been a opener.

[@mohsenghafari7652]

hi. please help me. how to create custom model from many pdfs in Persian language? tank you.

[@goodstuff5666]

Very nice tutorial! Could you guys share the slides? Thanks.