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

uv sync
trio login
swanlab login

Evaluation uses TRIO's OpenAI-compatible API. If the configuration file does not yet exist, run:

cp 02-opd/.env.example 02-opd/.env

Then fill in PYTRIO_API_KEY in 02-opd/.env.

1. Download the data

uv run python 02-opd/00-download-dataset.py

This prepares the Medical SFT, MedQA-zh and C-Eval data.

2. Start training

First train the Medical SFT teacher:

uv run python 02-opd/02-medical-sft.py \
    --num-epochs 3 \
    --batch-size 16 \
    --max-length 2048 \
    --swanlab-mode online

Record the SFT sampler_weights path printed at the end, then start Medical OPD:

uv run python 02-opd/03-medical-opd-async.py \
    --teacher-model-path 'trio://RUN_ID/sampler_weights/SFT_WEIGHTS_NAME' \
    --steps 300 \
    --batch-size 4 \
    --group-size 4 \
    --sample-size 0 \
    --max-tokens 2048 \
    --learning-rate 4e-5 \
    --save-every-steps 300 \
    --swanlab-mode online

Replace RUN_ID and SFT_WEIGHTS_NAME with the actual values from the previous run.

3. Run evaluation

Replace the model path below with the sampler weights path printed by OPD training:

uv run python 02-opd/01-eval-medical.py \
    --model 'trio://RUN_ID/sampler_weights/OPD_WEIGHTS_NAME' \
    --max-tokens 1024 \
    --concurrency 16

uv run python 02-opd/01-eval-ceval.py \
    --model 'trio://RUN_ID/sampler_weights/OPD_WEIGHTS_NAME' \
    --max-tokens 8192 \
    --concurrency 16

Per-question results and aggregate metrics are written to 02-opd/eval-results/.