harindhar10/chemberta4-evals

chemberta4 evaluations

Raw generations and scores from evaluating open-source LLMs, served by vLLM, on MoleculeNet property prediction and USPTO-50k single-step retrosynthesis.

Each run lives under <task>/<run_id>/:

• results.jsonl / generations.jsonl: one record per request, holding the model's raw output, any reasoning trace, and the parsed prediction.
• metrics.csv / summary.json: the scores.
• run_config.json: the evaluation hyperparameters.

Each run is also a dataset config:

from datasets import load_dataset
ds = load_dataset("harindhar10/chemberta4-evals", "<task>--<run_id>")

Runs

retrosynthesis / uspto_50k_test_250_Qwen3.8-27B_autosolve_USPTO

Config: retrosynthesis--uspto_50k_test_250_Qwen3.8-27B_autosolve_USPTO. Files: generations.jsonl, summary.json, summary.txt, run_config.json

Hyperparameters

hyperparameter	value
task	retrosynthesis
model	hosted_vllm/Qwen/Qwen3.8-27B
temperature	pipeline schedule
mode	autosolve
prompt	deepretro
data	uspto_50k_test_250
stability_check	True
hallucination_check	True
az_model	USPTO
top_k	3
thinking	False
max_output_tokens	1024
workers	64
git_commit	062ed83b7ac5594a88b2508cd79f1e7d36250578
timestamp	2026-11-11T16:56:55+00:00

view	all correct %	any correct %	MaxFrag %
top-1	24.8	34.4	28.8
top-k	25.2	35.2	30

molecules	parse failures %	invalid SMILES %	mean proposals
250	3.2	0	1.484

chemllmbench / Qwen3.8-27B

Config: chemllmbench--Qwen3.8-27B. Files: results.jsonl, metrics.csv, summary.json, run_config.json

Hyperparameters

hyperparameter	value
task	chemllmbench
model	hosted_vllm/Qwen/Qwen3.8-27B
temperature	0
subtasks	smiles_to_iupac, iupac_to_smiles, smiles_to_formula, iupac_to_formula, bbbp, bace, tox21, hiv, clintox, reaction_prediction, reactant_selection, ligand_selection, solvent_selection, retrosynthesis, mol_caption, mol_design, Buchwald-Hartwig, Suzuki-Miyaura
shots	0
strategy	random
thinking	False
max_output_tokens	1024
workers	64
git_commit	a7d8534df8fac7514a9ad72ca682d5898bd3c5d6
timestamp	2026-11-13T17:46:20+00:00

task	k	strategy	metric	value	n	parse fail rate
smiles_to_iupac	0	random	accuracy	0	100	0
iupac_to_smiles	0	random	accuracy	0.04	100	0
smiles_to_formula	0	random	accuracy	0.07	100	0
iupac_to_formula	0	random	accuracy	0.11	100	0
bbbp	0	random	accuracy	0.51	100	0
bace	0	random	accuracy	0.62	100	0
tox21	0	random	accuracy	0.58	100	0
hiv	0	random	accuracy	0.87	100	0
clintox	0	random	accuracy	0.95	100	0
reaction_prediction	0	random	top1_accuracy	0	100	0
reactant_selection	0	none	accuracy	0.47	100	0
ligand_selection	0	none	accuracy	0.07	100	0
solvent_selection	0	none	accuracy	0.32	100	0
retrosynthesis	0	random	top1_accuracy	0	100	0
Buchwald-Hartwig	0	random	accuracy	0.21	100	0
Suzuki-Miyaura	0	random	accuracy	0.36	100	0