Skip to content

GETTING_STARTED.md

Latest commit

 

History

History
118 lines (80 loc) · 2.78 KB

GETTING_STARTED.md

File metadata and controls

118 lines (80 loc) · 2.78 KB

Getting Started

This guide is for users who want to install the repository, prepare local data, and run the main paper-facing entrypoints without first reading the full implementation audit.

1. Install dependencies

Recommended baseline:

python -m pip install -U pip
python -m pip install -r requirements.txt --no-build-isolation
python -m pip check

Notes:

  • Python 3.11 is the reference version.
  • The repository is Linux-first and expects a CUDA-compatible PyTorch stack for the full training path.
  • Reproduce scripts export the repo root on PYTHONPATH automatically.

For lightweight local development and CI-style checks, editable install is also supported:

python -m pip install -e .[test]

2. Prepare local datasets

The public repository does not ship prepared datasets. Generate them locally:

bash scripts/data/prepare_all.sh --out_dir data

Strict verification:

bash scripts/data/prepare_all.sh --out_dir data --strict 1

See DATA_SOURCES.md for provenance and SHA256 pinning.

3. Run a wiring smoke test

This checks task loading, prompt rendering, and evaluator wiring. It is not a model-quality regression test.

bash scripts/reproduce/smoke.sh --task math --limit 1 --print_example 0

You can also run:

bash scripts/reproduce/smoke.sh --task code --limit 1 --print_example 0

4. Run paper-facing workflows

SFT-only eval sweep

bash scripts/reproduce/paper_main_results.sh sweep \
  --registry configs/main_results_registry.yaml \
  --only_methods SFT

Full training matrix

export PRETRAIN='Qwen/Qwen2.5-3B-Instruct'
bash scripts/reproduce/paper_train.sh

Full main-results sweep

bash scripts/reproduce/paper_main_results.sh sweep \
  --registry configs/main_results_registry.yaml

Paper analyses

bash scripts/reproduce/paper_analysis_figs.sh fig2 \
  --suite math \
  --run_c3 ckpt/_runs/<C3_run_dir> \
  --run_mappo ckpt/_runs/<MAPPO_run_dir> \
  --run_magrpo ckpt/_runs/<MAGRPO_run_dir> \
  --run_sft ckpt/_runs/_sft_main_results/<SFT_dir> \
  --mappo_critic_ckpt <PATH_TO_MAPPO_CRITIC>

5. Understand the implementation

6. Release hygiene

Before publishing the repository, run:

bash scripts/audit/pre_release.sh

Single-command preflight:

bash scripts/reproduce/preflight_repro.sh --task math

The release surface must not include local generated directories such as data/, artifacts/, ckpt/, runs/, wandb/, or models/. See RELEASE_POLICY.md.

For the full local release gate, use:

bash scripts/audit/release_gate.sh