The official implementation of ICSE 2026 paper: FoundRoot: Towards Foundation Model for Root Cause Analysis via Structured Deep Thinking.

You can download the model checkpoints from huggingface: FoundRoot-14B.

• The preprocessing pipeline that converts the raw DejaVu datasets (A1/A2, https://doi.org/10.5281/zenodo.6955909) into the evaluation datasets A and B is provided in utils/preprocess_dejavu.py.

(Update) Model checkpoints have been uploaded!

• Install python>3.11. We recommend using Linux environment with 8 x Nvidia A100 GPUs for training and evaluation.
• Run pip install -r requirements.txt.

• We have provided all the training and evaluation datasets under the data/ folder (datasets A - J). Datasets A - D are the evaluation datasets used in the paper (63 / 90 / 10 / 255 cases respectively) and contain only test.jsonl; datasets E - J are the training datasets with train.jsonl and test.jsonl splits.
• Data Augmentation: You need to set the SOURCE_DATASET_DIR, OUTPUT_DATASET_DIR and path to a local LLM PATH_TO_LLM in utils/augment_training_datasets.py first. Please refer to the content under datasets/ folder to prepare your own datasets. Then use python3 -m utils.augment_training_datasets to augment the datasets.
• Genenerate WarmUp-SFT Datasets: You need to set LLM_MODEL_PATH in utils/generate_warmup_dataset.py first. Then use python3 -m utils.generate_warmup_dataset -d [PATH_TO_INPUT_DATA_DIR] to generate the dataset.

• You can use LLaMA-Factory for WarmUp SFT Training.
• You can use trl for RL training with the provided receipe: train/rl_receipe.yaml.

Download the model checkpoints from huggingface: FoundRoot-14B.

We have provided the evaluation code for both local LLMs and APIs:

• Local LLMs: Set the LLM model path in evaluation/evaluate_local_llms.py. Then run python3 -m evaluation.evaluate_local_llms to get the evalution results.
• LLM APIs: Set the model name, base_url and api_key in evaluation/evaluate_api.py. Then run python3 -m evaluation.evaluate_api to get the evaluation results.

If you discover a potential security issue, please contact ByteDance Security via the security center or email [email protected]. Do not open public GitHub issues for vulnerabilities.

This project is licensed under the MIT License (see LICENSE).

• Qwen (https://github.com/QwenLM/Qwen2.5)
• DeepSpeed (https://www.deepspeed.ai/)
• vLLM (https://github.com/vllm-project/vllm)
• Flash-Attention (https://github.com/Dao-AILab/flash-attention)