Makefile

.PHONY: setup smoke collect build-data distill train-s1 calibrate eval eval-v2 eval-25 eval-cpu fetch-games agent-local push-hf test clean

PY ?= python3.13

setup:            ## create venv and install deps (Python 3.12+ required)
	$(PY) -m venv .venv
	.venv/bin/pip install -U pip
	.venv/bin/pip install -r requirements.txt
	@echo ">> done. activate with: source .venv/bin/activate"

smoke:            ## FULL local bring-up on the built-in mock game (no GPU, no downloads)
	$(PY) data/collect_selfplay.py --config configs/collect.yaml
	$(PY) data/build_dataset.py --in data/traces --out data/dataset
	$(PY) train/distill_alignment.py --config configs/train.yaml
	$(PY) train/train_system1.py --config configs/train.yaml
	$(PY) train/calibrate_rlcd.py --config configs/train.yaml
	$(PY) train/evaluate_local.py --config configs/train.yaml --games mock-pushbox

collect:          ## self-play trace collection (mock game or public games)
	$(PY) data/collect_selfplay.py --config configs/collect.yaml

build-data:       ## traces -> decision-question training samples
	$(PY) data/build_dataset.py --in data/traces --out data/dataset

distill:          ## stage A: align student (image+text) states with the frozen teacher's vision tower
	$(PY) train/distill_alignment.py --config configs/train.yaml

train-s1:         ## stage B: train System-1 decision heads (choice CE, NanoJev recipe)
	$(PY) train/train_system1.py --config configs/train.yaml

calibrate:        ## stage C: RLCD calibration + temperature fitting (absorbed from Laya)
	$(PY) train/calibrate_rlcd.py --config configs/train.yaml

eval:             ## evaluate the hybrid agent (mock game by default)
	$(PY) train/evaluate_local.py --config configs/train.yaml --games mock-pushbox

eval-v2:          ## one-command eval of the released v2 checkpoint (auto backend; see ops/run_eval.py)
	$(PY) ops/run_eval.py --checkpoint hf:tchbcb/qwenjev@t4_assets_v2/checkpoints --games mock

eval-25:          ## evaluate on the 25 public games (needs arc-agi + environment_files, see train.md §4.1)
	$(PY) ops/run_eval.py --checkpoint hf:tchbcb/qwenjev@t4_assets_v2/checkpoints --games all25

eval-cpu:         ## CPU harness test (tiny backend, no checkpoint, 60 actions)
	$(PY) ops/run_eval.py --backend tiny --no-checkpoint --games mock --max-actions 60

fetch-games:      ## cache all 25 public games locally (anonymous arcade key, ~3MB, no Kaggle account)
	$(PY) ops/fetch_games.py

agent-local:      ## run the hybrid agent locally via the ARC-AGI-3-Agents runner
	$(PY) ../ARC-AGI-3-Agents/main.py --agent=qwenjevagent --game=ls20

push-hf:          ## push this repo to HuggingFace (HF_TOKEN from env or .secrets)
	$(PY) scripts/push_hf.py

test:             ## unit tests (CPU-only, no downloads)
	$(PY) -m pytest tests -q

clean:
	rm -rf __pycache__ .pytest_cache build data/traces data/dataset eval_report.json