Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
Direct Preference Optimization
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
Henry Cui
February 24, 2024
Science
480
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Direct Preference Optimization
Henry Cui
February 24, 2024
More Decks by Henry Cui
See All by Henry Cui
プロダクション言語モデルの情報を盗む攻撃 / Stealing Part of a Production Language Model
zchenry
1
260
Diffusion Model with Perceptual Loss
zchenry
0
540
レンズの下のLLM / LLM under the Lens
zchenry
0
250
Go with the Prompt Flow
zchenry
0
260
Mojo Dojo
zchenry
0
280
ことのはの力で画像の異常検知 / Anomaly Detection by Language
zchenry
0
760
驚愕の事実!LangChainが抱える問題 / Problems of LangChain
zchenry
0
340
MLOps初心者がMLflowを触る / MLflow Brief Introduction
zchenry
0
230
{{guidance}}のガイダンス / Guidance of guidance
zchenry
0
210
Other Decks in Science
See All in Science
[TMLR 2026, Featured Certification] Double Bounded α-Divergence Optimization for Density Estimation
gkazunii
1
120
生成AIが科学とRAにもたらしていること:メタサイエンスの視点から
rmaruy
0
140
Conversation is the New Dashboard: 属人性を排除する第4世代BIツールの勢力図
shomaekawa
1
680
生成AI・プレプリント時代における 研究成果公開の再設計 ― トップカンファレンス文化はどこへ向かうのか / Redesigning the Dissemination of Research Outputs in the Age of Generative AI and Preprints — Where Is the Top-Conference Culture Heading?
ykiyota
0
30k
科学で迫る勝敗の法則-スポーツデータ分析の最前線 (刈谷市連携講座.2026年7月) / The principle of victory discovered by science. at Kariya City, 2027.07
konakalab
0
170
バランスって大事だね
akasan
1
150
Note agro-climatique et prairies - Numéro 7
institudelelevage
PRO
0
270
Where does the rigor go? Research software and the future of trustworthy science.
arfon
0
270
CVPR2026_VGGTとその仲間たち
mickey_0226
0
1.1k
[NLP2026 参加報告会] AI for Science まとめ / NLP2026
lychee1223
0
2k
Utiliser Bitcoin sans Internet
rlifchitz
0
380
20260410_SystemsThinking
takusamar
1
160
Featured
See All Featured
We Have a Design System, Now What?
morganepeng
55
8.3k
Bootstrapping a Software Product
garrettdimon
PRO
306
120k
Git: the NoSQL Database
bkeepers
PRO
432
67k
Conquering PDFs: document understanding beyond plain text
inesmontani
PRO
4
3.1k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
360
31k
A brief & incomplete history of UX Design for the World Wide Web: 1989–2019
jct
2
510
The B2B funnel & how to create a winning content strategy
katarinadahlin
PRO
1
540
Measuring & Analyzing Core Web Vitals
bluesmoon
9
1k
Game over? The fight for quality and originality in the time of robots
wayneb77
1
290
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
1
610
The Organizational Zoo: Understanding Human Behavior Agility Through Metaphoric Constructive Conversations (based on the works of Arthur Shelley, Ph.D)
kimpetersen
PRO
0
460
Amusing Abliteration
ianozsvald
1
310
Transcript
Direct Preference Optimization 機械学習の社会実装勉強会第32回 Henry 2024/2/24
内容 ▪ NeurIPS 2023 Outstanding Main Track Runner-Ups 受賞 ▪
著者に有名な先生が多い 2
モチベーション ▪ 大量テキストで学習した言語モデルを望ましい挙動に微調整 する必要(Alignment) • 大量コードの平均能力でなく、少量存在の優れたコードに • 一般大衆のもつ誤認識でなく、それを修正すべき ▪ Alignmentを達成するために、現状2段階の複雑な強化学習
手法を使うので、それと理論上等価なシンプルな手法を提案 3
RLHFアプローチの3ステップ ▪ SFT: Supervised fine-tuning ▪ Rewardモデルを学習する • RewardモデルがBradley-Terry (BT)に従う想定
• BTの仮定で導出する損失関数 ▪ RL Fine-tune • Rewardモデルを使って、下記損失関数でfine-tune ▪ 提案法はRewardとRL Fine-tuneをまとめて、rewardモデルを 使わずに学習 4
提案法DPO ▪ RL Fine-tuneの損失関数の最適解 ▪ 上記最適解をrewardモデルを取り出すよう書き換える • Your Language Model
Is Secretly a Reward Model ▪ Rewardモデルを学習する損失関数に代入する • BTモデルのお陰で、Zが消える • Directに言語モデルを最適化できるようになる 5
実験 ▪ 3つのタスクで評価 • controlled sentiment generation • summarization •
single-turn dialogue ▪ 複数スケールのデータセットでRHLFと同等またはそれ以上の 性能を確認 ▪ 多数のオープンソース言語モデルに実装 6