iLOVE2D/wetlabintheloop

Wet Lab in the Loop

Can a language model tell, before an experiment is run, whether it will actually help, where "actually" is decided by wet-lab measurements the model never sees? This dataset tests that question at two scales.

1. Dry loop (1,372 decisions). An automated agent improved molecular-property models on 22 TDC ADMET benchmarks (66 searches, about 21 experiments each). The agent kept or discarded each change by its validation score. We relabel every decision by what happened on the held-out test split, which consists of assay measurements. Validation and test disagree often. Of the changes the validation score accepted, 40% made the test metric worse. Of the changes it rejected, 38% would have improved it.
2. Wet loop (63 cases). A single real antibacterial plate: 28 compounds, 4 species, 2 replicates, OD600 endpoint. From it we compute experiment-design questions with exact answers: which species is uninterpretable, how much replication is needed, whether the purchasing model's ranking predicted the plate (it did not), and so on.

All splits are by benchmark. The 4 test benchmarks (CYP2C9_Veith, Clearance_Hepatocyte_AZ, DILI, VDss_Lombardo) never appear in training, so memorising one search's code lineage cannot transfer across the split.

Configs

config	split sizes	input → target	what it measures
judge	1,123 / 249	task card + history of val scores + proposed change + code diff → yes/no	pre-run: will this change raise the test metric?
posthoc	1,123 / 249	judge prompt + the run's own validation delta → yes/no	does the validation gain transfer to test? baseline = trust validation
proposal_sft	515 / 113	state → one-line proposal that improved test	supervised data for a proposal model
preference_pairs	267 / 68	state, chosen (test ↑), rejected (test ↓)	DPO-style data; pairs come from parallel searches at the same budget position
pairwise_choice	— / 67	state + options A/B → letter	pick the next experiment; derived from preference_pairs/test, order randomised
wetlab_cases	— / 63	plate readings + question → ANSWER: line	wet-lab experiment design with exact, deterministic answers

Fields

judge / posthoc: benchmark, model, exp_id, prompt, label, harness_status (keep|discard|crash), val_delta, test_delta. val_delta and test_delta are changes vs. the parent experiment. label = yes iff test_delta > 0.

wetlab_cases: id, family, prompt, gold (a JSON-encoded string; decode with json.loads), grader, plus trap for the cross-species items and sigma for the sample-size items.

wet-lab family	n	answer type	what a correct answer requires
qc_exclude	1	set	dropping the species whose growth control did not grow (OD 0.18)
replicate_rank	1	ranking (Kendall)	ranking species by replicate agreement
inhibition_calc	30	number, ±1 pt	normalising to the growth control
cross_species	15	species code	comparing normalised inhibition; every item is a trap where the raw-OD comparison gives the wrong answer
hit_count	6	integer, exact	applying a threshold across a full 28-compound table
replicate_outlier	3	compound id	finding the worst replicate pair to re-test
sample_size	3	integer, ±1	a power calculation from the measured replicate noise
scorer_transfer	2	rho ±0.05 + yes/no	checking whether the purchasing model's scores predicted the plate
mic_design	1	MC	recognising a censored MIC
next_round	1	MC	spending the next round on measuring the proxy's transfer, not harvesting its top

Evaluation protocol

pip install "vllm>=0.10" huggingface_hub   # B200 / sm_100 needs torch>=2.7 cu128
huggingface-cli download iLOVE2D/wetlabintheloop --repo-type dataset --local-dir wlitl
cd wlitl
python eval/run_eval.py --model Qwen/Qwen2.5-7B-Instruct --out runs/qwen2.5-7b
python eval/run_eval.py --model Qwen/Qwen2.5-7B-Instruct \
    --lora results/qwen2.5-7b-instruct_lora_judge/adapter --out runs/qwen2.5-7b-lora
python eval/run_eval.py --model Qwen/Qwen3-32B --tp 2 --enable-thinking true \
    --max-new-tokens 8192 --out runs/qwen3-32b
python eval/run_eval.py --backend openai --base-url $URL --model $NAME --out runs/$NAME
python eval/summarize.py runs          # leaderboard.md + leaderboard.csv
sbatch eval/run_b200.sh                # the full model sweep on one 8x B200 node

Rules

• Greedy decoding (temperature 0). Report the exact model, the chat template, and whether thinking mode was on.
• Never train on any test split. Model selection uses benchmarks from the train split only (the reference adapter used 3 of the training benchmarks as a dev set).
• judge, posthoc and pairwise_choice are scored by the first-token log-odds (yes vs no, A vs B), giving AUROC. Reasoning models and API models instead generate a reply, which is parsed, so they report accuracy only.
• wetlab_cases is graded deterministically from the final ANSWER: line (eval/grading.py). No LLM judge is involved, and a missing answer line scores 0.
• Report 95% CIs: run_eval.py writes bootstrap CIs for AUROC and Wilson CIs for pairwise accuracy. With 249 / 67 / 63 items, differences under ~0.05 are usually noise.

Primary metrics

task	metric	baselines that must be beaten
judge	AUROC; precision and test gain of the kept set	always-no; the validation rule
judge + validation rule	precision of "keep only if both agree"	validation rule alone
posthoc	accuracy and AUROC	the validation rule itself (val_delta > 0)
pairwise_choice	accuracy	0.5
wetlab_cases	macro score over families; trap score on cross_species	0 for raw-OD heuristics on traps

Reference results

Single A40 smoke run, vLLM 0.8.5, greedy decoding, --max-new-tokens 2048. Per-item outputs are in results/reference_runs/.

model	judge AUROC (95% CI)	judge + val rule: precision of kept	posthoc AUROC (95% CI)	pairwise acc	wetlab macro	wetlab micro
always "no"	0.500	—	0.500	0.500	—	—
validation rule (val_delta > 0)	0.544	0.510 (n=96)	0.544	—	—	—
Qwen2.5-7B-Instruct	0.550 (0.48–0.62)	0.750 (n=20)	0.560 (0.49–0.63)	0.493	0.407	0.444
+ LoRA judge adapter (this repo)	0.565 (0.49–0.63)	0.711 (n=38)	0.614 (0.54–0.68)	0.507	0.467	0.365

How to read this

• Neither model beats chance with confidence on the pre-run judge task: both CIs include 0.5. The LoRA adapter's gain over its base model is +0.015 AUROC, which is not significant.
• The LoRA adapter does help on posthoc (0.614, CI excludes 0.5), and on the "keep only if the validation rule and the model both agree" filter: precision 0.71 vs 0.51 for the validation rule alone, with roughly the same total test gain. (The base model's 0.75 rests on only 20 kept items.)
• pairwise_choice is at chance for both models, with a strong bias towards answering A (67–72% of answers).
• On wetlab_cases, the 7B models:
  • mostly get the multiple-choice design questions right;
  • fail qc_exclude (both say no species should be excluded) and scorer_transfer;
  • make arithmetic slips on inhibition_calc (for example, averaging the growth control wrongly).
  • The adapter was trained on dry-lab decisions only, and its wet-lab score does not improve.

Provenance and limitations

• Dry loop. The benchmarks are from the Therapeutics Data Commons ADMET group, using TDC's scaffold splits. The underlying datasets carry their original licences. The searches, code diffs and relabelling are ours.
• Preference pairs contrast actions taken by different parallel searches at the same benchmark and budget position. The states are comparable but not identical, so they are not counterfactuals from one state.
• Wet loop. The compounds were supplied by a collaborator and are released only as anonymous ids (C01–C28): no structures, vial ids or sources. The compounds were not designed by our agent. The plate is one round, with one concentration, and A. baumannii failed to grow.
• The prompts show validation scores only, as the original agent saw them. All labels come from the test split.
• Two definitions of "the validation rule was right" appear. dataset_meta.json reports 0.683 agreement on test, measured against the running best. The judge labels are measured against the parent experiment, where the same rule scores 0.554.