FMs in Bedrock - Overview
Ref: https://www.udemy.com/course/aws-ai-practitioner-certified/learn/lecture/44886343
- Choosing an adequate FM depends on a lot of factors!
- Model types, performance requirements, capabilities, constraints, compliance
- Level of customization, model size, inference options, licensing agreements, context windows, latency
- Multimodal models (varied types of input and outputs)
- Multimodal models can do e.g. both text and image
- 💡 Smaller models: more cost-effective & more limited in outputs
- Amazon Titan
- High-performing FM from AWS
- Image, text, multimodal model choices via fully-managed APIs
- Can be customized with your own data
Comparison table of important FMs
| Bedrock FM |
Features |
Use cases |
| Amazon Titan |
- High-performance text model |
|
- Lowest price
- Multi-language | Content creation, classification, education… |
| Meta LLaMa | -Â Large-scale tasks
- Dialogue
- Only English | Text generation, customer service… |
| Anthropic Claude | - High-capacity text generation- Highest input tokens
- Multi-language | Analysis, forecasting, document comparison… |
| Stability.ai Stable Diffusion | - Image generation | Image creation for advertising, media... |
- Stable Diffusion used mainly for image generation
- Cheapness: Titan > Llama > Claude
- Input tokens: Claude > Titan > Llama
- i.e. how big/long can the prompt be
- ‼️ Remember features and use cases!!
- 💡 The FMs are slowly converging to the same thing, but right now there are differences between them
DEMO
Ref: https://www.udemy.com/course/aws-ai-practitioner-certified/learn/lecture/44886347
- Can compare outputs of different FMs for same prompts
- Custom Models → Influence an FM with your own data:
- Fine-tuning job: applied once
- Continued Pre-training job: applied continuously
- Fine-tuning an FM
- Uses S3 → Bedrock needs a service role with S3 permissions
- Training data
- Validation data (if existent)
- Output data
- Hyperparameters: present in ML training jobs
- configurations on how the ML training algorithm should behave
- very useful for data scientists
- âť—Â Must purchase provisioned throughput to run a fine-tuned model
- Very expensive!!
- 💡 You don't need to purchase Provisioned Throughput to actually DO the fine-tuning job (training), but you will have to purchase Provisioned Throughput to run the fine-tuned model (inference)