Run all commands from the repository root. The project requires Python >=3.13.

uv sync
trio login
swanlab login

1. Download the data

Prepare the training set and AIME25 evaluation set:

uv run python 05-retool/prepare_data.py
uv run python 04-opsd/00-datasets.py --only aime25

2. Start training

Begin with a 20-step validation run:

uv run python 05-retool/train.py \
    --max-steps 20 \
    --save-every 5 \
    --run-name retool-qwen35-4b-step20

For the full experiment, set --max-steps to 200 and --save-every to 50.

3. Run evaluation

First evaluate the base model:

uv run python 05-retool/eval.py \
    --mode retool \
    --val-n 12 \
    --temperature 1.0 \
    --top-p 0.7 \
    --output 05-retool/eval-results/aime25-retool-base.jsonl

Then evaluate the trained sampler weights:

uv run python 05-retool/eval.py \
    --mode retool \
    --val-n 12 \
    --temperature 1.0 \
    --top-p 0.7 \
    --model-path 'trio://RUN_ID/sampler_weights/RETOOL_WEIGHTS_NAME' \
    --output 05-retool/eval-results/aime25-retool-step200.jsonl