Trustworthy AI

This repository is a collection of trustworthy AI related works from Huawei Noah's Ark Lab.

---

gCastle

• A causal structure learning toolchain containing various functionalities related to causal learning and evaluation. A tech report describing the toolbox is available here.
• The package offers a number of causal discovery algorithms, most of which are gradient-based, hence the name: gradient-based Causal structure learning pipeline.

Competition

• Information and baselines for causality-related competitions arranged by Noah's Ark Lab.
• Previous competitions were held at PCIC 2021, PCIC 2022, and NeurIPS 2023.

Datasets

• Real-world datasets released by Huawei Noah's Ark Lab.
• Code for generating various synthetic datasets.

Research

• Research works related to causality. We will continuously add new methods here.
• Currently contains implementations of CausalVAE, GAE, and causal discovery with reinforcement learning.

---

TabbyTSFM

• Tabby, a time series foundation model proposed by the Paris team of Huawei Noah's Ark Lab, supporting forecasting, classification, and anomaly detection. Its pretraining code is fully open-sourced here.
• Data processing: TabbyTSFM/src/tabby/data holds the GIFT-Eval/BLAST shard reader, the KernelSynth Arrow reader, and the online CauKer V2 generator that samples series from structural causal models; shared input preprocessing is in TabbyTSFM/src/tabby/utils.
• Pretraining: TabbyTSFM/recipes/pretrain contains the distributed mixed-data trainer and the train_165k.sh release entry point.
• Tests and evaluation: TabbyTSFM/tests covers the release contract, checkpoint loading across every accepted layout, and the GIFT-Eval script end to end; TabbyTSFM/benchmarks holds the benchmark harnesses for GIFT-Eval and TIME (forecasting), UCR (classification), and TSB-AD (anomaly detection).
• The weights are published on Hugging Face at paris-noah/Tabby