Dataset Source Code
Steering LLMs for better instruction following
This repository contains the code for the paper “Improving Instruction-Following in Language Models through Activation Steering,” presented at ICLR 2025.
Researcher Tools: code, datasets, & models
Discover an index of datasets, SDKs, APIs and open-source tools developed by Microsoft researchers and shared with the global academic community below. These experimental technologies—available through Azure AI Foundry Labs (opens in new tab)—offer a glimpse into the future of AI innovation.
Dataset Source Code
SeCom
Repository for “On Memory Construction and Retrieval for Personalized Conversational Agents”
Dataset Source Code
CFPO: Content-Format Integrated Prompt Optimization
CFPO (Content-Format Integrated Prompt Optimization) is a novel methodology that concurrently optimizes both prompt content and format for Large Language Models (LLMs) through an iterative refinement process. It addresses the limitations of existing prompt optimization…
Dataset Source Code
TimeCraft
Time series generation technology plays a vital role in alleviating data scarcity, especially in scenarios where collecting real-world data is expensive, time-consuming, or impractical. It also enables privacy-preserving analysis by producing realistic but non-identifiable synthetic…
Dataset Source Code
RAS: Region-Adaptive Sampling for Diffusion Transformers
The RAS library is an open-source implementation of Regional-Adaptive Sampling (RAS), a novel diffusion model sampling strategy that introduces regional variability in sampling steps. Unlike conventional methods that uniformly process all image regions, RAS dynamically…
Dataset Source Code
Distilled Decoding
This is the official repository of paper “Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching.” We propose Distilled Decoding (DD) to distill a pre-trained image auto-regressive model to few steps for…
Dataset Source Code
Skill Slice Insights
This is the official code repository for the paper “Unearthing Skill-level Insights for Understanding Tradeoffs of Foundation Models”. All rationales, localized skills, and skill-slices for the 12 datasets studied in the paper can also be accessed…
Dataset Source Code
Belief State Transformer
This is the official codebase for the paper “The Belief State Transformer”, based on the nanoGPT repository by Andrej Karpathy.