MACHINE LEARNING

→ View original post on X — @berkeley_ai

UC Berkeley AI Ranking Project Becomes Industry’s Leading System Benchmark

By

@berkeley_ai

–

07 December 2024 2h16

Work led by BAIR students @ml_angelopoulos and @infwinston has "…quickly become the most-watched ranking of the world’s best AI systems." https://
wsj.com/tech/ai/the-uc
-berkeley-project-that-is-the-ai-industrys-obsession-bc68b3e3
… @Cal_Engineer @UCBerkeley @Berkeley_EECS

SFT and RL Fine-Tuning: Complementary Approaches for Model Optimization

By

–

07 December 2024 2h08

I think that SFT will remain useful for fine-tuning a model to a new specific task, and RL fine-tuning becomes an interesting additional toolset further to push the model toward a desired kind of answer while also allowing it more flexibility in the way it “thinks” of the answer

OpenAI’s RLFT Boosts o1-mini Accuracy in Law, Finance, Engineering

By

–

07 December 2024 2h08

OpenAI sees huge potential in domains like Law, Finance, and Engineering, where “objectively correct” answers matter. Early results are promising: RLFT helped o1-mini boost accuracy in a gene example from 25% to 31%!

Question-Answer Pairs Training Without Perfect Ground Truth

By

–

07 December 2024 2h08

Why is this interesting? It lets us use question-answer pairs without needing the full chain-of-thought input or even a perfect ground-truth output for SFT. Graders can assign nuanced scores (0–1), improving model outputs incrementally—even when answers are partially

RLFT Evaluates Complete Answers Instead of Individual Tokens

By

–

07 December 2024 2h08

Instead of evaluating each token, one at a time, RLFT grades entire answers—perfect for scenarios like OpenAI’s o1, where we can’t access intermediate tokens before the “end_of_thought.”

OpenAI Releases Reinforcement Fine-Tuning for Flexible Model Training

By

–

07 December 2024 2h08

Today, we got Reinforcement Fine-Tuning (RLFT)! (OpenAI releases day 2) Unlike Supervised Fine-Tuning (SFT) (already available on OpenAI for models like 4o), RLFT trains models with a flexible, non-fixed objective.

→ View original post on X — @jeande_d

Research Visit at CMU: Multilingual Multimodal AI Focus

By

@jeande_d

–

07 December 2024 1h22

Some life updates: starting a research visit at @LTIatCMU
, working with truly amazing @gneubig and the team on range of topics in multimodal AI, most notably multilingual multimodal LLMs and multimodal evaluation. Excited for the journey and the opportunity to work along