Alexhf825/AutoMR-pangu

AutoMR for Pangu

This repository runs AutoMR with a local openPangu-Embedded-7B model on Huawei Ascend.

Project Structure

.
├── MATH500.sh
├── automr
├── checkpoints
│   └── MATH
├── embedder_server.sh
├── generator_server.sh
├── main.py
├── math_train.sh
├── openPangu-Embedded-7B
└── processed_data
    ├── MATH
    └── MATH_500

1. Prepare the Repository

hf download Alexhf825/AutoMR-pangu --local-dir AutoMR-pangu
cd AutoMR-pangu

2. Download the Model into the Project

All scripts now expect the model path to be ./openPangu-Embedded-7B.

hf download FreedomIntelligence/openPangu-Embedded-7B --local-dir ./openPangu-Embedded-7B

Before starting vLLM, edit ./openPangu-Embedded-7B/config.json and change:

"max_position_embeddings": 131072

Without this change, vLLM will not support the required 128k context length.

3. Dataset

processed_data will be committed together with the repository, so no separate dataset download step is needed.

The repository uses:

• processed_data/MATH/train.jsonl
• processed_data/MATH/val.jsonl
• processed_data/MATH_500/test.jsonl

4. Start the Servers

Run the embedder and generator in two separate terminals:

bash embedder_server.sh

bash generator_server.sh

5. Run Evaluation

The evaluation script is MATH500.sh:

bash MATH500.sh

By default it evaluates on processed_data/MATH_500/test.jsonl and uses the checkpoint under checkpoints/MATH/pangu/token_budget_128000/.

6. Run Training

bash math_train.sh

This trains on processed_data/MATH with token_budget=128000 and saves checkpoints under checkpoints/MATH/pangu/token_budget_128000/.