task_id stringclasses 6
values | task_type stringclasses 2
values | question stringclasses 6
values | inputs dict | outputs listlengths 0 5 | success_rules unknown | reference unknown | options dict |
|---|---|---|---|---|---|---|---|
study1/q1 | oe | Compare metabolite abundances between gastric cancer (GC) and non-GC participants in Cohort 1. Identify differential features and report effect sizes and statistical evidence. | {
"files": [
{
"path": "study/study1/code/GC_diagnosis_model/Data/discovery_set.xlsx",
"description": "426 Cohort-1 participants, 147 row-sum normalized features plus Batch/sample_id/type(N=NGC,P=GC)/batch. This contains both supplied training and internal-test participants, so it is not an additional... | [
{
"path": "analysis.py or analysis.R",
"type": "script"
},
{
"path": "results.json",
"type": "json"
},
{
"path": "differential.csv",
"type": "table",
"description": "One row per analyzed measured feature, with its identifier, defined effect measure, raw p-value and multiplicity-a... | {
"checks": [
{
"id": "delivery",
"category": "evidence",
"description": "Provide a final quantitative answer, reproducible analysis code, results.json and the feature-level differential table. Core values must agree between code, table and answer; no optional figure, extra exploration or exact ... | {
"answer": "The reference comparison identified 45 differential features: 20 higher and 25 lower in GC. The leading higher feature was 4-Hydroxyphenylpyruvic acid_neg-013 (FC 1.86,FDR 6.4e-17), and the leading lower feature was Succinate_neg-079 (FC 0.66,FDR 1.5e-33).",
"entry": "benchmark/study1/reference/run.py"... | null |
study1/c1 | mcq | [Single-answer] In Cohort 1, compare gastric cancer (GC, type P) with non-GC (type N) using the 147 row-sum normalized metabolite features. Apply a two-sided asymptotic Mann-Whitney U test with tie and continuity correction, Benjamini-Hochberg correction across all 147 tests, and arithmetic-mean fold change GC/NGC. Def... | {
"files": [
{
"path": "study/study1/code/GC_diagnosis_model/Data/discovery_set.xlsx",
"description": "426 Cohort-1 participants, 147 row-sum normalized features plus Batch/sample_id/type(N=NGC,P=GC)/batch. This contains both supplied training and internal-test participants, so it is not an additional... | [] | {} | {
"answer": "A",
"entry": "benchmark/study1/reference/run.py",
"result": "benchmark/study1/reference/results.json",
"analysis": "differential",
"result_ids": [
"differential"
]
} | {
"A": "20 higher and 25 lower features.",
"B": "25 higher and 20 lower features.",
"C": "22 higher and 23 lower features.",
"D": "45 higher and no lower features."
} |
study1/q2 | oe | Using the supplied metabolomics data, identify a parsimonious feature combination and develop a model to distinguish GC from non-GC. Assess discrimination and external-validation performance, reporting the features used and quantitative results. | {
"files": [
{
"path": "study/study1/code/GC_diagnosis_model/Data/discovery_set.xlsx",
"description": "426 Cohort-1 participants, 147 row-sum normalized features plus Batch/sample_id/type(N=NGC,P=GC)/batch. This contains both supplied training and internal-test participants, so it is not an additional... | [
{
"path": "analysis.py or analysis.R",
"type": "script"
},
{
"path": "results.json",
"type": "json"
},
{
"path": "predictions.csv",
"type": "table",
"description": "Participant-level predictions for the evaluated internal and external cohorts, retaining actual participant IDs and... | {
"checks": [
{
"id": "delivery",
"category": "evidence",
"description": "Submit the final answer, reproducible model code, results.json and participant-level predictions identifying evaluation cohort, true label and score. Provide the selected features and external results; optional ROC figure ... | {
"answer": "The reference Random Forest used ten features: Succinate_neg-079, Uridine_neg-088, S-Adenosyl-methionine_pos-139, N-Acetyl-D-glucosamine 6-phosphate_neg-061, Serotonin_pos-142, Pyroglutamic acid_neg-072, Neopterin_pos-117, Lactic acid_neg-055, 2-Aminooctanoic acid_pos-006, NMN_pos-162. Its internal AUC/s... | null |
study1/c2 | mcq | [Single-answer] Using the diagnosis outcome tables and their matched participant IDs, how does the observed proportion of GC cases differ across the supplied training, internal-test and external-test cohorts? Count state=1 as GC and retain the supplied partitions. | {
"files": [
{
"path": "study/study1/code/GC_diagnosis_model/Data/discovery_set.xlsx",
"description": "426 Cohort-1 participants, 147 row-sum normalized features plus Batch/sample_id/type(N=NGC,P=GC)/batch. This contains both supplied training and internal-test participants, so it is not an additional... | [] | {} | {
"answer": "D",
"entry": "benchmark/study1/reference/run.py",
"result": "benchmark/study1/reference/results.json",
"analysis": "diagnosis",
"result_ids": [
"diagnosis"
]
} | {
"A": "All three cohorts contain approximately one-third GC cases.",
"B": "The training and internal-test cohorts contain approximately two-thirds GC cases, while the external cohort contains one-third.",
"C": "The external cohort has a lower GC proportion than either training or internal testing.",
"D": "The ... |
study1/q3 | oe | Using the supplied GC metabolomics and follow-up data, develop an overall-survival prediction model. Account for right censoring and report predictive performance and its uncertainty. | {
"files": [
{
"path": "study/study1/code/GC_prognosis_model/Data/Xtrain.xlsx",
"description": "121 training GC participants; Sample is patient ID, remaining 147 columns are row-sum normalized features."
},
{
"path": "study/study1/code/GC_prognosis_model/Data/ytrain.xlsx",
"descrip... | [
{
"path": "analysis.py or analysis.R",
"type": "script"
},
{
"path": "results.json",
"type": "json",
"schema": {
"answer": "Conclusion supported by the computed analysis",
"metrics": {
"named_metric": "computed numerical value"
},
"uncertainty": {
"met... | {
"checks": [
{
"id": "delivery",
"category": "evidence",
"description": "Provide the final survival conclusion, self-contained code, results.json, predictions and train/test memberships for the primary evaluation. Submitted principal numbers must be regenerated. Auxiliary training CV need not i... | {
"answer": "The 28-feature Random Survival Forest achieved a test Harrell C-index of 0.826087 (features selected using the same test outcomes), with a fixed-model, fixed-panel conditional 95% bootstrap interval of [0.675028,0.938462].",
"entry": "benchmark/study1/reference/run.py",
"result": "benchmark/study1/re... | null |
study1/c3 | mcq | [Single-answer] In the supplied overall-survival training and test cohorts, compare the fractions of patients with an observed death during their recorded follow-up. OS_State=1 denotes death, OS_State=0 denotes right censoring, and OS is follow-up time in months. Which statement agrees with the data and this endpoint d... | {
"files": [
{
"path": "study/study1/code/GC_prognosis_model/Data/Xtrain.xlsx",
"description": "121 training GC participants; Sample is patient ID, remaining 147 columns are row-sum normalized features."
},
{
"path": "study/study1/code/GC_prognosis_model/Data/ytrain.xlsx",
"descrip... | [] | {} | {
"answer": "B",
"entry": "benchmark/study1/reference/run.py",
"result": "benchmark/study1/reference/results.json",
"analysis": "prognosis",
"result_ids": [
"survival"
]
} | {
"A": "The training cohort has 97 observed deaths and the test cohort has 50 observed deaths.",
"B": "The training cohort has 24/121 observed deaths (about 19.8%) and the test cohort has 10/60 (about 16.7%); these are observed event fractions, not mortality probabilities at a common fixed time.",
"C": "Both coho... |
MetaboBench
Evaluate agents on metabolomics analysis tasks. Each row is one task; the Viewer reads the original benchmark/studyN/tasks.json directly.
| Study | Open tasks | Multiple-choice tasks |
|---|---|---|
| study1 | 3 | 3 |
Columns: task_id, task_type, question, inputs, outputs, success_rules, reference; multiple-choice tasks also use options.
Open tasks require an independent analysis script and results.json plus declared artifacts. SR requires effective execution and all necessary success rules to pass; VER measures valid clean execution. Report resource use separately. Multiple-choice accuracy is separate.
import json
from pathlib import Path
from huggingface_hub import snapshot_download
root = Path(snapshot_download("sanjarmarmar/metabobench", repo_type="dataset", local_dir="metabobench"))
tasks = json.loads((root / "benchmark/study1/tasks.json").read_text())
input_files = [root / f["path"] for f in tasks[0]["inputs"]["files"]]
All input and reference paths are relative to the downloaded dataset root, with no parent-directory traversal or author-specific absolute paths. load_dataset loads task rows only; snapshot_download also downloads the actual data, reference programs and evaluation code.
To run the provided agents and scorer, use Linux, Python 3.11 and bubblewrap, then configure your own model endpoint:
cd metabobench
python3.11 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
cp metabobench/.env.example metabobench/.env.llm
# Edit .env.llm to point to your own model service.
bash run.sh check --live
bash run.sh agent --dataset benchmark/study1/tasks.json --q-type oe --output-dir outputs/sdk_qwen/study1/oe
bash run.sh score --dataset benchmark/study1/tasks.json --q-type oe --results outputs/sdk_qwen/study1/oe/answers.json
Prepared tasks use the active Python environment. Docker is needed only to reproduce the legacy study1 author reference:
docker build -f docker/study1.Dockerfile -t metabobench-study1:latest .
Inputs and provenance are under study/studyN/; reference programs, results and tables are under benchmark/studyN/reference/. Reference answers and author materials are public: agents should receive only the question, declared inputs and public requirements.
Source methodology limitations remain documented in each study's context and reference results. Third-party materials retain their original sources and terms; the project MIT license does not relicense all input data. No agent outputs, model credentials, build records or paper full texts are included.
Project code: GitHub.
- Downloads last month
- 119