Makefile

.PHONY: help setup test validate preflight data sft dpo rl graphs all clean
.DEFAULT_GOAL := help

PY      ?= .venv/bin/python
HUPD    ?= hupd/sample/2016
AUG     ?= hupd/aug2014
TOKENIZER ?= tokenizer_qwen35.json

help:  ## Show this help
	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) \
	  | awk 'BEGIN{FS=":.*?## "}{printf "  \033[36m%-14s\033[0m %s\n", $$1, $$2}'

setup:  ## Create .venv (Python 3.12) and install dependencies
	uv venv --python 3.12 .venv
	uv pip install -r requirements.txt
	@test -f .env || (cp .env.example .env && echo "Created .env -- add your TINKER_API_KEY")

test:  ## Run the offline test suite (no API key, no data rebuild needed)
	$(PY) tests/test_rewards.py
	$(PY) tests/test_dataset.py

validate:  ## Structural checks on the built dataset
	$(PY) scripts/validate_dataset.py data/sft

preflight:  ## Verify everything end to end, including the Tinker API
	$(PY) scripts/preflight.py

calibrate:  ## Re-check the reward function against real granted claims
	$(PY) claim_drafter/rewards.py data/sft/train.jsonl 2000

# ---------------------------------------------------------------- data
# The built datasets are ~600MB and are NOT committed. manifests/ holds the
# provenance (patent numbers) so these rebuild byte-identically.

hupd:  ## Download the HUPD Jan-2016 sample (~370MB)
	mkdir -p hupd
	curl -L -o hupd/sample-jan-2016.tar.gz \
	  https://huggingface.co/datasets/HUPD/hupd/resolve/main/data/sample-jan-2016.tar.gz
	tar -xzf hupd/sample-jan-2016.tar.gz -C hupd

tokenizer:  ## Download the Qwen3.5 tokenizer (exact token counts)
	curl -L -o $(TOKENIZER) \
	  https://huggingface.co/Qwen/Qwen3.5-9B/resolve/main/tokenizer.json

data: hupd tokenizer  ## Rebuild the SFT dataset end to end (~1 hour, mostly fetching)
	$(PY) scripts/select_targets.py --hupd $(HUPD) --output data/cache/targets_granted.txt
	$(PY) scripts/fetch_claims.py --kind granted \
	  --input data/cache/targets_granted.txt --output data/cache/granted_claims.jsonl
	$(PY) scripts/augment_thin_domains.py 2014 $(AUG)
	$(PY) scripts/select_targets.py --hupd $(AUG) --output data/cache/targets_aug.txt \
	  --domains biotech_life_sciences,energy_environment,pharma_medical_devices
	$(PY) scripts/fetch_claims.py --kind granted \
	  --input data/cache/targets_aug.txt --output data/cache/granted_claims.jsonl
	$(PY) scripts/build_sft_dataset.py $(TOKENIZER) --sources $(HUPD) $(AUG) --out-dir data/sft
	$(PY) scripts/validate_dataset.py data/sft

dpo-data:  ## Build the DPO preference pairs (needs data/sft first)
	$(PY) scripts/select_targets.py --hupd $(HUPD) --output data/cache/targets_pubs.tsv \
	  --publications --manifest data/sft/manifest.jsonl
	$(PY) scripts/fetch_claims.py --kind filed \
	  --input data/cache/targets_pubs.tsv --output data/cache/filed_claims.jsonl
	$(PY) scripts/build_dpo_pairs.py $(TOKENIZER) --out data/dpo

# ---------------------------------------------------------------- training
sft:  ## Stage 1: supervised fine-tuning
	$(PY) training/train_sft.py --epochs 2

dpo:  ## Stage 2: DPO. CKPT defaults to the SFT run's final checkpoint
	$(PY) training/train_dpo.py --from-checkpoint \
	  $(or $(CKPT),$(shell $(PY) scripts/last_checkpoint.py runs/sft))

rl:  ## Stage 3: GRPO. CKPT defaults to the DPO run's final checkpoint
	$(PY) training/train_rl.py --from-checkpoint \
	  $(or $(CKPT),$(shell $(PY) scripts/last_checkpoint.py runs/dpo))

# ---------------------------------------------------------------- reporting
graphs:  ## Build charts + summary from every stage's logs (run after `all`)
	$(PY) scripts/plot_runs.py --runs runs --out runs/graphs

all:  ## Run the whole pipeline end to end, then draw the graphs
	$(MAKE) sft
	$(MAKE) dpo
	$(MAKE) rl
	$(MAKE) graphs

clean:  ## Remove generated data and run artifacts (keeps manifests/ and samples/)
	rm -rf data/sft data/dpo data/cache runs export __pycache__ \
	       claim_drafter/__pycache__ scripts/__pycache__ tests/__pycache__