From 32e3bc88a29961e90f3d722483c07285fb7c040a Mon Sep 17 00:00:00 2001 From: Michael Sawitzki Date: Mon, 13 May 2024 13:31:59 +0200 Subject: [PATCH] Update Chapter 07/08 --- content/chapters/07_gpt/07_01_gpt.md | 2 +- content/chapters/07_gpt/07_02_gpt2.md | 1 + content/chapters/07_gpt/07_04_tasks.md | 4 +++- content/chapters/07_gpt/07_05_discussion.md | 8 +++++--- content/chapters/08_llm/08_01_instruction_tuning.md | 13 ++++++++++--- content/chapters/08_llm/08_02_cot.md | 13 ++++++++++--- content/chapters/08_llm/08_03_emerging.md | 12 +++++++++--- content/chapters/08_llm/_index.md | 2 +- 8 files changed, 40 insertions(+), 15 deletions(-) diff --git a/content/chapters/07_gpt/07_01_gpt.md b/content/chapters/07_gpt/07_01_gpt.md index e42a763..2d5cc8f 100644 --- a/content/chapters/07_gpt/07_01_gpt.md +++ b/content/chapters/07_gpt/07_01_gpt.md @@ -8,7 +8,7 @@ By fine-tuning the model on task-specific data with minimal changes to the archi -### Lecture slides +### Lecture Slides {{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter07-gpt/slides-71-gpt.pdf" >}} diff --git a/content/chapters/07_gpt/07_02_gpt2.md b/content/chapters/07_gpt/07_02_gpt2.md index 79fd55b..68c99e3 100644 --- a/content/chapters/07_gpt/07_02_gpt2.md +++ b/content/chapters/07_gpt/07_02_gpt2.md @@ -6,6 +6,7 @@ GPT-2 [1] builds upon its predecessor with a larger model size, more training da +### Lecture Slides {{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter07-gpt/slides-72-gpt2.pdf" >}} diff --git a/content/chapters/07_gpt/07_04_tasks.md b/content/chapters/07_gpt/07_04_tasks.md index 5a52185..c23f2d4 100644 --- a/content/chapters/07_gpt/07_04_tasks.md +++ b/content/chapters/07_gpt/07_04_tasks.md @@ -1,5 +1,5 @@ --- -title: "Chapter 7.4: Tasks & Performance" +title: "Chapter 07.04: Tasks & Performance" weight: 7004 --- @@ -7,4 +7,6 @@ GPT-3 has X-shot abilities, meaning it is able to perform tasks with minimal or +### Lecture Slides + {{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter07-gpt/slides-74-tasks.pdf" >}} diff --git a/content/chapters/07_gpt/07_05_discussion.md b/content/chapters/07_gpt/07_05_discussion.md index c2ad037..f8878f1 100644 --- a/content/chapters/07_gpt/07_05_discussion.md +++ b/content/chapters/07_gpt/07_05_discussion.md @@ -1,12 +1,14 @@ --- -title: "Chapter 7.5: Discussion: Ethics and Cost" +title: "Chapter 07.05: Discussion: Ethics and Cost" weight: 7005 --- -In discussing GPT-3's ethical implications, it's crucial to consider its potential societal impact, including issues surrounding bias, misinformation, and data privacy. With its vast language generation capabilities, GPT-3 has the potential to disseminate misinformation at scale, posing risks to public trust and safety. Additionally, the model's reliance on large-scale pretraining data raises concerns about reinforcing existing biases present in the data, perpetuating societal inequalities. Furthermore, the use of GPT-3 in sensitive applications such as content generation, automated customer service, and decision-making systems raises questions about accountability, transparency, and unintended consequences. As such, responsible deployment of GPT-3 requires careful consideration of ethical guidelines, regulatory frameworks, and robust mitigation strategies to address these challenges and ensure the model's ethical use in society. +In discussing GPT-3's ethical implications, it is crucial to consider its potential societal impact, including issues surrounding bias, misinformation, and data privacy. With its vast language generation capabilities, GPT-3 has the potential to disseminate misinformation at scale, posing risks to public trust and safety. Additionally, the model's reliance on large-scale pretraining data raises concerns about reinforcing existing biases present in the data, perpetuating societal inequalities. Furthermore, the use of GPT-3 in sensitive applications such as content generation, automated customer service, and decision-making systems raises questions about accountability, transparency, and unintended consequences. As such, responsible deployment of GPT-3 requires careful consideration of ethical guidelines, regulatory frameworks, and robust mitigation strategies to address these challenges and ensure the model's ethical use in society. -{{< video id="TfrSKiOecWI" >}} + + +### Lecture Slides {{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter07-gpt/slides-75-discussion.pdf" >}} diff --git a/content/chapters/08_llm/08_01_instruction_tuning.md b/content/chapters/08_llm/08_01_instruction_tuning.md index de4ef3c..4e0a75d 100644 --- a/content/chapters/08_llm/08_01_instruction_tuning.md +++ b/content/chapters/08_llm/08_01_instruction_tuning.md @@ -1,10 +1,17 @@ --- -title: "Chapter 8.1: Instruction Fine-Tuning" +title: "Chapter 08.01: Instruction Fine-Tuning" weight: 8001 --- -In this chapter we introduce instruction-tuning, which is a technique that allows us to adapt the models to follow instructions. +Instruction fine-tuning aims to enhance the adaptability of large language models (LLMs) by providing explicit instructions or task descriptions, enabling more precise control over model behavior and adaptation to diverse contexts. +This approach involves fine-tuning LLMs on task-specific instructions or prompts, guiding the model to generate outputs that align with the given instructions. By conditioning the model on explicit instructions, instruction fine-tuning facilitates more accurate and tailored responses, making LLMs more versatile and effective in various applications such as language translation, text summarization, and question answering. -{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter8-multilinguality/slides-81-why_multilingual.pdf" >}} +### Lecture Slides + +{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter08-llm/slides-81-instruction-tuning.pdf" >}} + +### Additional Resources + +- [Blog about Instruction Fine-Tuning](https://heidloff.net/article/instruct-tuning-large-language-models/) diff --git a/content/chapters/08_llm/08_02_cot.md b/content/chapters/08_llm/08_02_cot.md index 08995c2..a82001f 100644 --- a/content/chapters/08_llm/08_02_cot.md +++ b/content/chapters/08_llm/08_02_cot.md @@ -1,9 +1,16 @@ --- -title: "Chapter 8.2: Chain-of-thought Prompting" +title: "Chapter 08.02: Chain-of-thought Prompting" weight: 8002 --- -In this session we cover Chain-of-thoght Prompting, which is a technique to improve the performance of models without requiring additional training. +Chain of thought (CoT) prompting [1] is a prompting method that encourage Large Language Models (LLMs) to explain their reasoning. This method contrasts with standard prompting by not only seeking an answer but also requiring the model to explain its steps to arrive at that answer. By guiding the model through a logical chain of thought, chain of thought prompting encourages the generation of more structured and cohesive text, enabling LLMs to produce more accurate and informative outputs across various tasks and domains. -{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter8-multilinguality/slides-82-multilingual-wordembs.pdf" >}} +### Lecture Slides + +{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter08-llm/slides-82-chain-of-thought.pdf" >}} + +### References + +- [1] [Wei et al., 2022](https://arxiv.org/abs/2201.11903) + diff --git a/content/chapters/08_llm/08_03_emerging.md b/content/chapters/08_llm/08_03_emerging.md index f81958d..f91b983 100644 --- a/content/chapters/08_llm/08_03_emerging.md +++ b/content/chapters/08_llm/08_03_emerging.md @@ -1,9 +1,15 @@ --- -title: "Chapter 8.3: Emergent Abilities" +title: "Chapter 08.03: Emergent Abilities" weight: 8003 --- -Various researchers have reported that LLMs seem to have emergent abilities. In this section we discuss the concept of emergence in LLMs. +Various researchers have reported that LLMs seem to have emergent abilities. These are sudden appearances of new abilities when Large Language Models (LLMs) are scaled up. In this section we introduce the concept of emergent abilities and discuss a potential counterargument for the concept of emergence. -{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter8-multilinguality/slides-83-multilingual-transformers.pdf" >}} +### Lecture Slides + +{{< pdfjs file="https://github.com/slds-lmu/lecture_dl4nlp/blob/main/slides/chapter08-llm/slides-83-emergent-abilities.pdf" >}} + +### Additional Resources + +- [Article: Large Language Models' Emergent Abilities Are a Mirage](https://www.wired.com/story/how-quickly-do-large-language-models-learn-unexpected-skills/) \ No newline at end of file diff --git a/content/chapters/08_llm/_index.md b/content/chapters/08_llm/_index.md index 1b3cb5a..b68e78e 100644 --- a/content/chapters/08_llm/_index.md +++ b/content/chapters/08_llm/_index.md @@ -2,4 +2,4 @@ title: "Chapter 8: Large Language Models (LLMs)" --- -Here we cover Large Language Models and concepts, such as Instruction Fine-Tuning and Chain-of-thought Prompting. +In this chapter we cover LLM concepts, such as Instruction Fine-Tuning, Chain-of-Thought prompting and discuss the possbility of emerging abilities of LLMs.