AI Wiki - User contributions [en]

AI Talk

2026-01-03T13:55:17Z

Whale:

== Seoul, the Republic of Korea ==
You need to use the Meetup app or its website to participate!

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312248145/

AAT-001 The Evolution of LLMs - https://docs.google.com/presentation/d/1foywsiogALeYuooJAIaIVTnl6BF2d4gsuu7Ny6O9-j8/edit?usp=sharing

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312325713/

AAT-002 Beyond the Prompt: The Hidden Architecture of Modern LLMs - https://docs.google.com/presentation/d/1twFJGSrawfwiFquHoWLcNep6bo_lvBrfnscyVU7EhJI/edit?usp=sharing

[[2026-01-10 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676231/

AAT-003 Intro to Diffusion Model - https://docs.google.com/presentation/d/1lLFhIWkwm6gr1cHAHLNPq-cBBby8JDzXqaRTTbCbKzM/edit?usp=sharing

[[2026-01-24 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676413/

[[2026-02-07 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676418/

== Another City ==

AI Talk

2026-01-03T13:54:14Z

Whale: /* Seoul, the Republic of Korea */

== Seoul, the Republic of Korea ==
You need to use the Meetup app or its website to participate!

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312248145/

AAT-001 The Evolution of LLMs - <nowiki>https://docs.google.com/presentation/d/1foywsiogALeYuooJAIaIVTnl6BF2d4gsuu7Ny6O9-j8/edit?usp=sharing</nowiki>

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312325713/

AAT-002 Beyond the Prompt: The Hidden Architecture of Modern LLMs - <nowiki>https://docs.google.com/presentation/d/1twFJGSrawfwiFquHoWLcNep6bo_lvBrfnscyVU7EhJI/edit?usp=sharing</nowiki>

[[2026-01-10 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676231/

AAT-003 Intro to Diffusion Model - <nowiki>https://docs.google.com/presentation/d/1lLFhIWkwm6gr1cHAHLNPq-cBBby8JDzXqaRTTbCbKzM/edit?usp=sharing</nowiki>

[[2026-01-24 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676413/

[[2026-02-07 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312676418/

== Another City ==

MediaWiki:Sidebar

2025-12-15T12:42:33Z

Whale:

AI Talk

2025-12-15T12:40:55Z

Whale: /* Seoul, the Republic of Korea */

== Seoul, the Republic of Korea ==
You need to use the Meetup app or its website to participate!

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312248145/

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station https://www.meetup.com/seoulshare/events/312325713/

[[2026-01-10 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2026-01-17 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2026-01-24 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

== Another City ==

Main Page

2025-12-15T12:26:51Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local [[AI Talk]] groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword in the search bar to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref></syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

* '''Using chatbots:''' Use an advance chatbot to generate the content automatically. A prompt similar to the one below typically accomplishes the task.<syntaxhighlight lang="html">
Write an article about {title} (in the context of AI) with citations that can work on a MediaWiki page. The format of each citation should be:
<ref>Publication, “Title”, accessed 2025‑12‑14, https://url-of-the-reference</ref>
At the end of the document, add the following, inserting a line break before each item:
== References ==
<references />
</syntaxhighlight>

== Getting started ==
*[[Large Language Models]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

Main Page

2025-12-15T12:24:46Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local [[AI Talk]] groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref></syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

* '''Using chatbots:''' Use an advance chatbot to generate the content automatically. A prompt similar to the one below typically accomplishes the task.<syntaxhighlight lang="html">
Write an article about {title} (in the context of AI) with citations that can work on a MediaWiki page. The format of each citation should be:
<ref>Publication, “Title”, accessed 2025‑12‑14, https://url-of-the-reference</ref>
At the end of the document, add the following, inserting a line break before each item:
== References ==
<references />
</syntaxhighlight>

== Getting started ==
*[[Large Language Models]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

MediaWiki:Sidebar

2025-12-15T12:09:09Z

Whale:

MediaWiki:Sidebar

2025-12-15T12:07:15Z

AI Talk

2025-12-15T12:00:37Z

Whale: Created page with "== Seoul, the Republic of Korea == 2025-12-06 Sat - 12:50 PM AB Cafe at the Gangnam Station 2025-12-06 Sat - 12:50 PM AB Cafe at the Gangnam Station 2026-01-10 Sat - 12:50 PM AB Cafe at the Gangnam Station 2026-01-17 Sat - 12:50 PM AB Cafe at the Gangnam Station 2026-01-24 Sat - 12:50 PM AB Cafe at the Gangnam Station == Another City =="

== Seoul, the Republic of Korea ==
[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2025-12-06 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2026-01-10 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2026-01-17 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

[[2026-01-24 Sat]] - 12:50 PM AB Cafe at the Gangnam Station

== Another City ==

Main Page

2025-12-15T11:57:23Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local [[AI Talk]] groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[Large Language Models]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

Main Page

2025-12-15T11:50:03Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local [[AI Adoption Talk]] groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[Large Language Models]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

Main Page

2025-12-15T07:13:19Z

Whale: /* Getting started */

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[Large Language Models]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

Large Language Models

2025-12-15T07:12:36Z

Whale: Created page with "A '''large language model''' ('''LLM''') is a language model trained with self-supervised machine learning on a vast amount of text, designed for natural language processing tasks, especially language generation. The largest and most capable LLMs are generative pre-trained transformers (GPTs) and provide the core capabilities of modern chatbots. LLMs can be fine-tuned for specific tasks or guided by prompt engineering. These models acquire predictive power regarding synt..."

A '''large language model''' ('''LLM''') is a language model trained with self-supervised machine learning on a vast amount of text, designed for natural language processing tasks, especially language generation. The largest and most capable LLMs are generative pre-trained transformers (GPTs) and provide the core capabilities of modern chatbots. LLMs can be fine-tuned for specific tasks or guided by prompt engineering. These models acquire predictive power regarding syntax, semantics, and ontologies inherent in human language corpora, but they also inherit inaccuracies and biases present in the data they are trained on.

They consist of billions to trillions of parameters and operate as general-purpose sequence models, generating, summarizing, translating, and reasoning over text. LLMs represent a significant new technology in their ability to generalize across tasks with minimal task-specific supervision, enabling capabilities like conversational agents, code generation, knowledge retrieval, and automated reasoning that previously required bespoke systems.

LLMs evolved from earlier statistical and recurrent neural network approaches to language modeling. The transformer architecture, introduced in 2017, replaced recurrence with self-attention, allowing efficient parallelization, longer context handling, and scalable training on unprecedented data volumes. This innovation enabled models like GPT, BERT, and their successors, which demonstrated emergent behaviors at scale, such as few-shot learning and compositional reasoning.

Reinforcement learning, particularly policy gradient algorithms, has been adapted to fine-tune LLMs for desired behaviors beyond raw next-token prediction. Reinforcement learning from human feedback (RLHF) applies these methods to optimize a policy, the LLM's output distribution, against reward signals derived from human or automated preference judgments. This has been critical for aligning model outputs with user expectations, improving factuality, reducing harmful responses, and enhancing task performance.

Benchmark evaluations for LLMs have evolved from narrow linguistic assessments toward comprehensive, multi-task evaluations measuring reasoning, factual accuracy, alignment, and safety. Hill climbing, iteratively optimizing models against benchmarks, has emerged as a dominant strategy, producing rapid incremental performance gains but raising concerns of overfitting to benchmarks rather than achieving genuine generalization or robust capability improvements. <ref>Wikipedia</ref>

== References ==
<references />

Instruction Tuning

2025-12-15T07:10:58Z

Whale:

'''Instruction tuning''' is a technique used in the training of [[Large Language Models]] (LLMs) to improve their ability to follow natural language instructions. While [[Pre-training|pre-training]] enables a model to predict the next token in a sequence based on vast amounts of text data, it does not inherently teach the model to act as a helpful assistant or adhere to specific user commands. Instruction tuning bridges this gap by [[Fine-tuning|fine-tuning]] the pre-trained model on a dataset of instruction-output pairs.<ref>IBM, "What Is Instruction Tuning?", accessed 2025-12-15, https://www.ibm.com/think/topics/instruction-tuning</ref>

== Overview ==
The primary goal of instruction tuning is to align the model's behavior with human intent. A standard pre-trained LLM might respond to the prompt "Explain the theory of relativity" by generating a continuation like "was proposed by Albert Einstein in 1905," rather than providing the explanation requested. By training the model on examples where the input is an instruction (e.g., "Summarize this text") and the output is the desired response, the model learns to interpret and execute the user's intent.<ref>GeeksforGeeks, "Instruction Tuning for Large Language Models", accessed 2025-12-15, https://www.geeksforgeeks.org/artificial-intelligence/instruction-tuning-for-large-language-models/</ref>

This process is often considered a critical step in "alignment," or ensuring AI systems behave in accordance with human values and expectations, and serves as a precursor to more advanced techniques like [[Reinforcement Learning from Human Feedback]] (RLHF).<ref>Ouyang, L., et al., "Training language models to follow instructions with human feedback", accessed 2025-12-15, https://arxiv.org/abs/2203.02155</ref>

== Methodology ==
Instruction tuning typically follows [[Supervised Learning|supervised learning]] paradigms. The process involves compiling a dataset where each example consists of:

An instruction: A natural language command describing the task (e.g., "Translate the following sentence into French").

An input (optional): The context or data to operate on (e.g., "The cat sat on the mat").

An output: The target response (e.g., "Le chat s'est assis sur le tapis").

Datasets
Early instruction tuning relied on datasets like FLAN (Finetuned Language Net), which aggregated existing [[Natural Language Processing]] (NLP) tasks—such as translation, summarization, and reading comprehension—and converted them into instruction formats.<ref>Wei, J., et al., "Finetuned Language Models Are Zero-Shot Learners", accessed 2025-12-15, https://research.google/pubs/finetuned-language-models-are-zero-shot-learners/</ref>

Later approaches, such as Stanford Alpaca, demonstrated that high-quality instruction data could be synthesized by prompting a stronger teacher model (like [[GPT-3]]) to generate diverse instruction-response pairs, significantly reducing the cost of data collection.<ref>Taori, R., et al., "Stanford Alpaca: An Instruction-following LLaMA Model", accessed 2025-12-15, https://github.com/tatsu-lab/stanford_alpaca</ref>

The "Less is More" Hypothesis
Research has suggested that the quantity of instruction data may be less important than its quality. The LIMA (Less Is More for Alignment) study proposed the "Superficial Alignment Hypothesis," suggesting that an LLM acquires most of its knowledge during pre-training. Consequently, instruction tuning serves mainly to teach the model the specific format or style of interaction, achievable with as few as 1,000 carefully curated examples.<ref>Zhou, C., et al., "LIMA: Less Is More for Alignment", accessed 2025-12-15, https://arxiv.org/abs/2305.11206</ref>

== Benefits ==

Zero-Shot Generalization: Instruction-tuned models show improved performance on tasks they were not explicitly trained on, as they learn the general concept of following instructions.<ref>Wei, J., et al., "Finetuned Language Models Are Zero-Shot Learners", accessed 2025-12-15, https://research.google/pubs/finetuned-language-models-are-zero-shot-learners/</ref>

Steerability: Users can direct the model's output style, tone, and format more effectively.

Efficiency: Compared to full model re-training, instruction tuning is computationally cheaper and can be applied to smaller models to achieve performance comparable to larger, non-tuned models.

== References ==
<references />

Main Page

2025-12-15T07:09:34Z

Whale: /* Getting started */

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM]]
*[[Pre-training]]
*[[Instruction Tuning]]
*[[Alignment (AI)]]
[[Category:Misc.]]

Alignment (AI)

2025-12-15T07:07:45Z

Whale:

'''AI Alignment''' refers to the process of directing [[Artificial Intelligence]] (AI) systems, particularly [[Large Language Models]] (LLMs), to act in accordance with human intent and ethical values. While the core objective of a pre-trained model is simply to predict the next token in a sequence based on statistical patterns, the goal of alignment is to ensure the resulting behavior is helpful, honest, and harmless.<ref>IBM, "What is AI alignment?", accessed 2025-12-15, https://www.ibm.com/topics/ai-alignment</ref>

Without alignment, LLMs may generate outputs that are factually incorrect (hallucinations), biased, toxic, or dangerous, even if those outputs are statistically probable completions of the input prompt.

== Core Components ==

Alignment is generally conceptualized around three main criteria, often referred to as the "HHH" framework:<ref>Askell, A., et al., "A General Language Assistant as a Laboratory for Alignment", accessed 2025-12-15, https://arxiv.org/abs/2112.00861</ref>

* Helpful: The model should attempt to perform the task specified by the user concisely and efficiently.
* Honest: The model should avoid fabricating information or misleading the user.
* Harmless: The model should not generate offensive, discriminatory, or dangerous content, even if explicitly asked to do so.

== Techniques ==

Several methodologies have been developed to align models post-pre-training.

=== Reinforcement Learning from Human Feedback (RLHF) ===
RLHF is currently the dominant technique for aligning state-of-the-art models. It involves a multi-step process:

* Supervised Fine-Tuning (SFT): The model is trained on a dataset of high-quality instruction-response pairs written by humans.
* Reward Modeling: The model generates multiple responses to a prompt, and human labelers rank them from best to worst. A separate "reward model" is trained to predict these human preferences.
* Reinforcement Learning: The language model is optimized against the reward model using algorithms like [[Proximal Policy Optimization]] (PPO), learning to generate outputs that maximize the predicted reward.<ref>OpenAI, "Aligning language models to follow instructions", accessed 2025-12-15, https://openai.com/research/instruction-following</ref>

=== Constitutional AI (RLAIF) ===
As models become more capable, relying solely on human feedback becomes difficult and expensive. Constitutional AI, or Reinforcement Learning from AI Feedback (RLAIF), proposes using the AI itself to guide alignment. The model critiques and revises its own responses based on a set of high-level principles or a "constitution" (e.g., "Do not support illegal acts"). This allows for alignment scaling with less direct human intervention.<ref>Anthropic, "Constitutional AI: Harmlessness from AI Feedback", accessed 2025-12-15, https://www.anthropic.com/research/constitutional-ai</ref>

== Challenges ==

=== The Alignment Tax ===
There is often a trade-off between the safety of a model and its capabilities, sometimes referred to as the "alignment tax." Heavily aligned models may refuse benign requests (false refusals) or become less creative due to strict safety filtering.<ref>TechTarget, "AI alignment", accessed 2025-12-15, https://www.techtarget.com/whatis/definition/AI-alignment</ref>

=== Jailbreaking ===
Despite alignment efforts, users often find "jailbreaks"—adversarial prompts designed to bypass safety filters (e.g., asking the model to roleplay as a villain). Robust alignment requires continuous "red teaming" to identify and patch these vulnerabilities.<ref>Wei, A., et al., "Jailbroken: How Does LLM Safety Training Fail?", accessed 2025-12-15, https://arxiv.org/abs/2307.02483</ref>

== References ==

<references />

Alignment (AI)

2025-12-15T07:01:45Z

Whale: Created page with "<nowiki>'''</nowiki>AI Alignment<nowiki>'''</nowiki> refers to the process of directing <nowiki>Artificial Intelligence</nowiki> (AI) systems, particularly <nowiki>Large Language Models</nowiki> (LLMs), to act in accordance with human intent and ethical values. While the core objective of a pre-trained model is simply to predict the next token in a sequence based on statistical patterns, the goal of alignment is to ensure the resulting behavior is helpful, honest..."

<nowiki>'''</nowiki>AI Alignment<nowiki>'''</nowiki> refers to the process of directing <nowiki>[[Artificial Intelligence]]</nowiki> (AI) systems, particularly <nowiki>[[Large Language Models]]</nowiki> (LLMs), to act in accordance with human intent and ethical values. While the core objective of a pre-trained model is simply to predict the next token in a sequence based on statistical patterns, the goal of alignment is to ensure the resulting behavior is helpful, honest, and harmless.<nowiki><ref>IBM, "What is AI alignment?", accessed 2025-12-15, https://www.ibm.com/topics/ai-alignment</ref></nowiki>

Without alignment, LLMs may generate outputs that are factually incorrect (hallucinations), biased, toxic, or dangerous, even if those outputs are statistically probable completions of the input prompt.

<nowiki>== Core Components ==</nowiki>

Alignment is generally conceptualized around three main criteria, often referred to as the "HHH" framework:<nowiki><ref>Askell, A., et al., "A General Language Assistant as a Laboratory for Alignment", accessed 2025-12-15, https://arxiv.org/abs/2112.00861</ref></nowiki>

* '''Helpful:''' The model should attempt to perform the task specified by the user concisely and efficiently.
* '''Honest:''' The model should avoid fabricating information or misleading the user.
* '''Harmless:''' The model should not generate offensive, discriminatory, or dangerous content, even if explicitly asked to do so.

<nowiki>== Techniques ==</nowiki>

Several methodologies have been developed to align models post-pre-training.

=== Reinforcement Learning from Human Feedback (RLHF) ===
RLHF is currently the dominant technique for aligning state-of-the-art models. It involves a multi-step process:

* Supervised Fine-Tuning (SFT): The model is trained on a dataset of high-quality instruction-response pairs written by humans.
* Reward Modeling: The model generates multiple responses to a prompt, and human labelers rank them from best to worst. A separate "reward model" is trained to predict these human preferences.
* Reinforcement Learning: The language model is optimized against the reward model using algorithms like <nowiki>[[Proximal Policy Optimization]]</nowiki> (PPO), learning to generate outputs that maximize the predicted reward.<nowiki><ref>OpenAI, "Aligning language models to follow instructions", accessed 2025-12-15, https://openai.com/research/instruction-following</ref></nowiki>

=== Constitutional AI (RLAIF) ===
As models become more capable, relying solely on human feedback becomes difficult and expensive. Constitutional AI, or Reinforcement Learning from AI Feedback (RLAIF), proposes using the AI itself to guide alignment. The model critiques and revises its own responses based on a set of high-level principles or a "constitution" (e.g., "Do not support illegal acts"). This allows for alignment scaling with less direct human intervention.<nowiki><ref>Anthropic, "Constitutional AI: Harmlessness from AI Feedback", accessed 2025-12-15, https://www.anthropic.com/research/constitutional-ai</ref></nowiki>

<nowiki>== Challenges ==</nowiki>

=== The Alignment Tax ===
There is often a trade-off between the safety of a model and its capabilities, sometimes referred to as the "alignment tax." Heavily aligned models may refuse benign requests (false refusals) or become less creative due to strict safety filtering.<nowiki><ref>TechTarget, "AI alignment", accessed 2025-12-15, https://www.techtarget.com/whatis/definition/AI-alignment</ref></nowiki>

=== Jailbreaking ===
Despite alignment efforts, users often find "jailbreaks"—adversarial prompts designed to bypass safety filters (e.g., asking the model to roleplay as a villain). Robust alignment requires continuous "red teaming" to identify and patch these vulnerabilities.<nowiki><ref>Wei, A., et al., "Jailbroken: How Does LLM Safety Training Fail?", accessed 2025-12-15, https://arxiv.org/abs/2307.02483</ref></nowiki>

<nowiki>== References ==</nowiki>

<nowiki><references /></nowiki>

Main Page

2025-12-15T06:55:12Z

Whale: /* Getting started */

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM]]
*[[Pre-training]]
*[[Instruction Tuning]]
[[Category:Misc.]]

Instruction Tuning

2025-12-15T06:53:54Z

Whale: Created page with "'''Instruction tuning''' is a technique used in the training of Large Language Models (LLMs) to improve their ability to follow natural language instructions. While pre-training enables a model to predict the next token in a sequence based on vast amounts of text data, it does not inherently teach the model to act as a helpful assistant or adhere to specific user commands. Instruction tuning bridges this gap by fine-tuning the pre-tra..."

'''Instruction tuning''' is a technique used in the training of Large Language Models ([[LLM]]s) to improve their ability to follow natural language instructions. While [[Pre-training|pre-training]] enables a model to predict the next token in a sequence based on vast amounts of text data, it does not inherently teach the model to act as a helpful assistant or adhere to specific user commands. Instruction tuning bridges this gap by [[Fine-tuning|fine-tuning]] the pre-trained model on a dataset of instruction-output pairs.<ref>IBM, "What Is Instruction Tuning?", accessed 2025-12-15, https://www.ibm.com/think/topics/instruction-tuning</ref>

== Overview ==
The primary goal of instruction tuning is to align the model's behavior with human intent. A standard pre-trained LLM might respond to the prompt "Explain the theory of relativity" by generating a continuation like "was proposed by Albert Einstein in 1905," rather than providing the explanation requested. By training the model on examples where the input is an instruction (e.g., "Summarize this text") and the output is the desired response, the model learns to interpret and execute the user's intent.<ref>GeeksforGeeks, "Instruction Tuning for Large Language Models", accessed 2025-12-15, https://www.geeksforgeeks.org/artificial-intelligence/instruction-tuning-for-large-language-models/</ref>

This process is often considered a critical step in "alignment," or ensuring AI systems behave in accordance with human values and expectations, and serves as a precursor to more advanced techniques like [[Reinforcement Learning from Human Feedback]] (RLHF).<ref>Ouyang, L., et al., "Training language models to follow instructions with human feedback", accessed 2025-12-15, https://arxiv.org/abs/2203.02155</ref>

== Methodology ==
Instruction tuning typically follows [[Supervised Learning|supervised learning]] paradigms. The process involves compiling a dataset where each example consists of:

An instruction: A natural language command describing the task (e.g., "Translate the following sentence into French").

An input (optional): The context or data to operate on (e.g., "The cat sat on the mat").

An output: The target response (e.g., "Le chat s'est assis sur le tapis").

Datasets
Early instruction tuning relied on datasets like FLAN (Finetuned Language Net), which aggregated existing [[Natural Language Processing]] (NLP) tasks—such as translation, summarization, and reading comprehension—and converted them into instruction formats.<ref>Wei, J., et al., "Finetuned Language Models Are Zero-Shot Learners", accessed 2025-12-15, https://research.google/pubs/finetuned-language-models-are-zero-shot-learners/</ref>

Later approaches, such as Stanford Alpaca, demonstrated that high-quality instruction data could be synthesized by prompting a stronger teacher model (like [[GPT-3]]) to generate diverse instruction-response pairs, significantly reducing the cost of data collection.<ref>Taori, R., et al., "Stanford Alpaca: An Instruction-following LLaMA Model", accessed 2025-12-15, https://github.com/tatsu-lab/stanford_alpaca</ref>

The "Less is More" Hypothesis
Research has suggested that the quantity of instruction data may be less important than its quality. The LIMA (Less Is More for Alignment) study proposed the "Superficial Alignment Hypothesis," suggesting that an LLM acquires most of its knowledge during pre-training. Consequently, instruction tuning serves mainly to teach the model the specific format or style of interaction, achievable with as few as 1,000 carefully curated examples.<ref>Zhou, C., et al., "LIMA: Less Is More for Alignment", accessed 2025-12-15, https://arxiv.org/abs/2305.11206</ref>

== Benefits ==

Zero-Shot Generalization: Instruction-tuned models show improved performance on tasks they were not explicitly trained on, as they learn the general concept of following instructions.<ref>Wei, J., et al., "Finetuned Language Models Are Zero-Shot Learners", accessed 2025-12-15, https://research.google/pubs/finetuned-language-models-are-zero-shot-learners/</ref>

Steerability: Users can direct the model's output style, tone, and format more effectively.

Efficiency: Compared to full model re-training, instruction tuning is computationally cheaper and can be applied to smaller models to achieve performance comparable to larger, non-tuned models.

== References ==
<references />

Main Page

2025-12-14T11:03:47Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop us a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM|LLM - AI Wiki]]
*[[Pre-training]]
[[Category:Misc.]]

Main Page

2025-12-14T10:48:55Z

Whale:

Welcome to AI Wiki. Your contribution is greatly appreciated.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop me a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM|LLM - AI Wiki]]
*[[Pre-training]]

Main Page

2025-12-14T10:44:18Z

Whale:

Welcome to AI Wiki. Your contribution is valued.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop me a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link a page on this website. (EG. <code><nowiki>[[Pre-training]]</nowiki></code> leads to [[Pre-training]])
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM|LLM - AI Wiki]]
*[[Pre-training]]

Main Page

2025-12-14T10:40:19Z

Whale:

Welcome to AI Wiki. Your contribution is valued.

Check out your local AI Adoption Talk groups. Currently, we are running a weekly meetup in Seoul. Let us know if you are interested in running a discussion group in your area. Drop me a line at d k l i m @ d o c e n t c o r p . c o m.

If you would like to contribute to this website, please enter your keyword to create a new page.

* '''Create a new page:'''
*# Type the desired page title in the search bar.
*# If it doesn’t exist, you’ll see a red link saying ''Create the page''.
*# Click it, add your content, and save.
* '''Linking pages:''' Use wiki syntax like <code><nowiki>[[Page Title]]</nowiki></code> to link internally.
* '''Categories:''' Add <code><nowiki>[[Category:YourCategory]]</nowiki></code> at the bottom of a page to organize entries.
* '''Enable references:''' This Wiki uses the <code><nowiki><ref></nowiki></code> and <code><nowiki><references /></nowiki></code> tags.
*# To cite something inline: <syntaxhighlight>
This is a statement that needs a source. <ref>Publication, "Title", accessed 2025-12-14, https://url-of-the-reference</ref>
</syntaxhighlight>
*# At the bottom of the page (or wherever you want the list of references to appear), add: <syntaxhighlight>
== References ==
<references />

</syntaxhighlight>

== Getting started ==
*[[LLM|LLM - AI Wiki]]
*[[Pre-training]]

Pre-training

2025-12-14T10:23:15Z

Whale:

Pretraining in AI is the initial phase of training a model on a large dataset to learn general patterns before fine-tuning it for specific tasks.

=== What is Pretraining? ===
Pretraining refers to the process of training a machine learning model on a large, diverse dataset before it is fine-tuned for a specific task. This phase is crucial as it equips the model with foundational knowledge, allowing it to learn general features and patterns that can be applied across various domains. For instance, a language model like GPT-4 is pretrained on vast amounts of text data to understand grammar, semantics, and context.

=== How Does Pretraining Work? ===

# '''Initial Training''': The model is exposed to extensive data, which can be done through unsupervised or supervised learning. During this phase, it learns to recognize patterns and relationships within the data.
# '''Transfer Learning''': The knowledge gained during pretraining can be transferred to different tasks, significantly reducing the amount of labeled data needed for fine-tuning.
# '''Fine-tuning''': After pretraining, the model is adjusted for specific tasks, optimizing its parameters to improve performance on those tasks.

=== Applications of Pretraining ===
Pretraining is widely used in various AI fields, including:

* '''Natural Language Processing (NLP)''': Models pretrained on large text corpora can quickly adapt to tasks like sentiment analysis or machine translation.
* '''Computer Vision''': Pretrained models can recognize general features in images, which can then be fine-tuned for specific image classification tasks.
* '''Speech Recognition''': Pretraining helps models understand general audio patterns, making them more effective when fine-tuned for specific speech tasks.

=== Benefits of Pretraining ===

* '''Efficiency''': Pretraining saves time and computational resources by allowing models to start with a strong foundational understanding rather than training from scratch.
* '''Improved Performance''': Models that undergo pretraining generally perform better on complex tasks due to their broader knowledge base.
* '''Reduced Data Requirements''': Pretraining lowers the need for large amounts of labeled data during the fine-tuning phase, which is particularly beneficial in domains where labeled datasets are scarce. In summary, pretraining is a vital step in AI development that enhances model performance and adaptability across various applications. It allows for more efficient training processes and better utilization of available data. <ref>Knapsack, "What is Pretraining and Post Training AI", accessed 2025-12-14, https://blog.knapsack.ai/what-is-pretraining-and-post-training-ai</ref><ref>GeeksforGeeks, "What is Pre-Training and its Objective", accessed 2025-12-14, https://www.geeksforgeeks.org/artificial-intelligence/what-is-pre-training-and-its-objective/</ref><ref>All About AI, "Pretraining", accessed 2025-12-14, https://www.allaboutai.com/ai-glossary/pretraining/</ref><ref>Baeldung, "What Does Pre-training a Neural Network Mean?", accessed 2025-12-14, https://www.baeldung.com/cs/neural-network-pre-training</ref><ref>TED San Francisco, "Pre-training", accessed 2025-12-14, https://tedai-sanfrancisco.ted.com/glossary/pre-training/</ref>
== References ==
<references />

Pre-training

2025-12-14T10:10:39Z

Whale: Created page with "Pretraining in AI is the initial phase of training a model on a large dataset to learn general patterns before fine-tuning it for specific tasks. === What is Pretraining? === Pretraining refers to the process of training a machine learning model on a large, diverse dataset before it is fine-tuned for a specific task. This phase is crucial as it equips the model with foundational knowledge, allowing it to learn general features and patterns that can be applied across var..."

Pretraining in AI is the initial phase of training a model on a large dataset to learn general patterns before fine-tuning it for specific tasks.

=== What is Pretraining? ===
Pretraining refers to the process of training a machine learning model on a large, diverse dataset before it is fine-tuned for a specific task. This phase is crucial as it equips the model with foundational knowledge, allowing it to learn general features and patterns that can be applied across various domains. For instance, a language model like GPT-4 is pretrained on vast amounts of text data to understand grammar, semantics, and context.

=== How Does Pretraining Work? ===

# '''Initial Training''': The model is exposed to extensive data, which can be done through unsupervised or supervised learning. During this phase, it learns to recognize patterns and relationships within the data.
# '''Transfer Learning''': The knowledge gained during pretraining can be transferred to different tasks, significantly reducing the amount of labeled data needed for fine-tuning.
# '''Fine-tuning''': After pretraining, the model is adjusted for specific tasks, optimizing its parameters to improve performance on those tasks.

=== Applications of Pretraining ===
Pretraining is widely used in various AI fields, including:

* '''Natural Language Processing (NLP)''': Models pretrained on large text corpora can quickly adapt to tasks like sentiment analysis or machine translation.
* '''Computer Vision''': Pretrained models can recognize general features in images, which can then be fine-tuned for specific image classification tasks.
* '''Speech Recognition''': Pretraining helps models understand general audio patterns, making them more effective when fine-tuned for specific speech tasks.

=== Benefits of Pretraining ===

* '''Efficiency''': Pretraining saves time and computational resources by allowing models to start with a strong foundational understanding rather than training from scratch.
* '''Improved Performance''': Models that undergo pretraining generally perform better on complex tasks due to their broader knowledge base.
* '''Reduced Data Requirements''': Pretraining lowers the need for large amounts of labeled data during the fine-tuning phase, which is particularly beneficial in domains where labeled datasets are scarce. In summary, pretraining is a vital step in AI development that enhances model performance and adaptability across various applications. It allows for more efficient training processes and better utilization of available data. <ref>{{Cite web|url=https://blog.knapsack.ai/what-is-pretraining-and-post-training-ai|title=Knapsack|access-date=2025-12-14}}</ref> <ref>{{Cite web|url=https://www.geeksforgeeks.org/artificial-intelligence/what-is-pre-training-and-its-objective/|title=Geek for Geeks|access-date=2025-12-14}}</ref> <ref>{{Cite web|url=https://www.allaboutai.com/ai-glossary/pretraining/|title=All About AI|access-date=2025-12-14}}</ref> <ref>{{Cite web|url=https://www.baeldung.com/cs/neural-network-pre-training|title=Baeldung|access-date=2025-12-14}}</ref> <ref>{{Cite web|url=https://tedai-sanfrancisco.ted.com/glossary/pre-training/|title=TED|access-date=2025-12-14}}</ref>

Main Page

2025-12-14T09:56:11Z

Whale:

Main Page

2025-12-14T09:47:28Z

Whale:

Main Page

2025-12-14T09:37:55Z

Whale: