feat: normalized [0,1] score by default + batch-first API (v0.1.1)

[kohbanye]

Makes the normalized [0, 1] score the default and reworks score() to be batch-first. Also trims the README. Bumps to v0.1.1.

Scoring API

• CLaMNPScorer.score(smiles, raw=False) now accepts a single SMILES or a list:
  • a list is scored in one batched forward pass (much faster than one-at-a-time) and returns a list;
  • a string returns a single score.
  • raw=True returns the raw log-likelihood ratio.
• This replaces score() / score_normalized() / batch_score() / batch_score_normalized() with one method.
• The default is the sigmoid-normalized score in [0, 1] (0.5 = neutral, higher = more natural-product-like), calibrated with sigmoid_k=2 to match the paper (DEFAULT_SIGMOID_K).

scorer = CLaMNPScorer.from_pretrained()
scorer.score(["CC(=O)Oc1ccccc1C(=O)O", "CN1C=NC2=C1C(=O)N(C(=O)N2C)C"])
# -> [0.44, 0.59]   (aspirin synthetic-like, caffeine natural-like)
scorer.score("CCO")             # single -> one score
scorer.score(["CCO"], raw=True) # raw log-likelihood ratio

CLI

• Default output is now the normalized [0, 1] score; --normalized is replaced by --raw-score for the raw ratio.
• Inputs are always scored in a single batch.

README

• Drop the "How it works" section; fold essentials into the intro (models are GPT-2, score is [0, 1]).
• Show real values (aspirin 0.44, caffeine 0.59) and the batch list-in/list-out form as primary.

Verification

• ruff check + ruff format --check: clean.
• ty check clamnp: clean (the overloaded score() adds no diagnostics).
• Functional test (no models): score("x") → float, score([...]) → list, raw=True returns raw, batch_score removed. Normalized values match sigmoid(2·raw) (0.4395 / 0.5859).

Numbers are computed from the published models' verified raw scores.

🤖 Generated with Claude Code

https://claude.ai/code/session_01C4cq7P9sF3VuHXFkDozKRs