Fine Tuning

Question 1

Distillation

What is Model Distillation?

Answer 1

Reduce a whole LLM to a much smaller, scoped-down model

Question 2

Distillation

Advantage to distilled models?

Answer 2

Smaller, faster, cheaper

Question 3

Distillation

How fast are distilled models?

Answer 3

In Amazon Bedrock are up to 500% faster

Question 4

Distillation

Accuracy loss with distilled models?

Answer 4

Less than 2% accuracy loss for use cases like RAG

Question 5

Fine Tuning

What is Fine-Tuning?

Answer 5

Changes the weights of base foundation model based on input data you give it.

Question 6

Fine Tuning

What format is input files to fine-tuning?

Answer 6

S3 JSON objects with keys like “prompt” and “completion”.

Question 7

Fine Tuning

Starting requirement of a model to use fine-tuning?

Answer 7

Must use Provisioned Throughput

Question 8

Fine Tuning

What is Instruction-based Fine Tuning?

Answer 8

Pass in prompts and expected completions

Question 9

Fine Tuning

Category of data passed into instruction-based fine-tuning?

Answer 9

Labeled data: has the expected completion included

Question 10

Fine Tuning

Why use instruction-based fine-tuning?

Answer 10

Domain-specific tasks where you supply the missing domain of data

Question 11

Fine Tuning

What is Continued Pre-training?

Answer 11

Pass large amt of unlabeled data, model trains on the whole thing

Question 12

Fine Tuning

What’s another name for Continued Pre-training?

Answer 12

Domain Adaptation Fine Tuning

Question 13

Fine Tuning

Example of continued pre-training?

Answer 13

Give a LLM the whole AWS documentation catalog

Question 14

Fine Tuning

What is Single-turn Messaging?

Answer 14

Type of instruction-based fine-tuning that mimics a chat bot

Question 15

Fine Tuning

Format for input to Single-turn Messaging?

Answer 15

Supply “user” (human question) and “assistant” (chatbot response), JSON

Question 16

Fine Tuning

What is Multi-turn Messaging?

Answer 16

Single-turn, but you supply lots of back-and-forth between human and chatbot

Question 17

Fine Tuning

Which is cheaper, instruction-based or continued pre-training?

Answer 17

Instruction-based: data is targeted and usually smaller

Question 18

Fine Tuning

Why is running a fine-tuned model is more expensive?

Answer 18

Have to use Provisioned Throughput

Question 19

Fine Tuning

What is Transfer Learning?

Answer 19

start with LLM or pre-trained model adapt it to a new related task

Question 20

Fine Tuning

how is transfer learning and fine-tuning related?

Answer 20

Fine Tuning is how you do it, Transfer Learning is the concept.