Rd kit deterministic

[Macaulay001]

Goal. Automatically verify the numeric / structural claims in each Q&A against RDKit ground truth derived from the SMILES. For claims like "degree of unsaturation = 3", "has 2 chiral centers", "contains a carboxylic acid group", RDKit can compute the right answer. We use this to audit label quality without an LLM judge.

[Copilot]

Pull request overview

Adds an RDKit-based deterministic auditing pipeline to validate numeric and structural claims in Chem2TextQA answers against SMILES-derived ground truth, producing per-compound and per-Q&A outputs without an LLM judge.

Changes:

• Introduces a topic bucketer to classify Q&A topics as structural vs functional.
• Implements an RDKit verifier that extracts checkable claims from answers and compares them to RDKit-derived facts, writing JSONL + Markdown summaries.
• Adds a script to emit per-Q&A JSONL records with aggregated RDKit labels.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
RDKit_Deterministic/scripts/topic_bucket.py	Adds shared helper to bucket freeform topic tags into structural/functional/other.
RDKit_Deterministic/scripts/rdkit_verifier.py	Adds RDKit ground-truth extraction + regex-based claim verification + reporting outputs.
RDKit_Deterministic/scripts/build_labeled_qa.py	Adds utility to join verifier output back onto each Q&A and compute an aggregate RDKit label.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

HEDGE_REGEX wraps symbol tokens like ~, ≈, ≥, ≤ in \b...\b, but word-boundaries don’t match around non-word characters. As a result, symbol-based hedges (e.g. ~3, ≤4) won’t be detected and will be treated as exact numeric claims. Consider handling these symbols without \b (e.g., separate alternation or lookarounds) so hedged claims are consistently skipped.

Suggested change

	# treated as a bound/approximation rather than a specific count.
	HEDGE_REGEX = re.compile(
	r"\b(?:fewer\s+than|less\s+than|more\s+than|greater\s+than|at\s+least|"
	r"at\s+most|up\s+to|under|over|about|approximately|roughly|around|"
	r"between|~|≈|≥|≤)\b",
	# treated as a bound/approximation rather than a specific count. Word/phrase
	# hedges use \b boundaries, but symbol hedges must be matched separately
	# because \b does not match around non-word characters like "~" or "≤".
	HEDGE_REGEX = re.compile(
	r"(?:\b(?:fewer\s+than|less\s+than|more\s+than|greater\s+than|"
	r"at\s+least|at\s+most|up\s+to|under|over|about|approximately|"
	r"roughly|around|between)\b|(?<!\w)(?:~|≈|≥|≤))",

[kushalviit]

@luistafoi if results are fine then comment "good to merge"