yamaa As a clinical trial statistician or programmer, challenge anything in this benchmark: describe a problem in Comments, or propose a fix with the Edit button on any source file.
- Task: Would a real study need this derivation, specified this way? Does it follow CDISC ADaM: variable names, controlled terms, record structure? Flag any rule that is ambiguous or that a sponsor would handle differently.
- Inputs: Do they look like data collected in a real trial, and do they include the situations you would expect to meet in one?
- Expected output: Would independent QC programming reproduce every value from these inputs and rules?
The other tabs beside Comments are background for assessment, not under review.
Summary
EditUnder reviewWhat this benchmark meansREADME.mdGoal: build the albumin change parameter _ALB, carrying each
visit's previous result (PREV_AVAL), the change from it (CHG),
and the previous visit's change (PREV2).
Input: laboratory (LB) records carrying a test code (LBTESTCD),
a numeric result (LBSTRESN), and a visit number (VISITNUM). Six
subjects: a four-visit chain that skips a visit number, a subject
whose middle visit has no result, a single-visit subject, a
two-visit subject, a subject with no change between two consecutive
visits, and one whose hemoglobin record produces no row.
Variables:
AVISITN: the analysis visit number, taken from the lab visit number.AVAL: the numeric albumin result for the visit.PREV_AVAL: theAVALof the visit just before inAVISITNorder, for the same subject; blank on the subject's first visit.CHG:AVAL - PREV_AVAL, the change since that visit; blank wheneverPREV_AVALis.PREV2: theCHGof the visit just before, read after that change is complete; blank until two earlier visits carry results.
Note: the previous visit's change is carried from its completed
value, so a record with no numeric result shortens the chain: it
gets no row, and the change and its lag compare against the latest
earlier visit with a result. CHG here is the change from the
previous visit, not the ADaMIG "change from baseline" definition; a
zero change is a real value (0), distinct from blank.
Inputs
Under review1 source fileinput/lb.csv
Edit15 rows| STUDYID | USUBJID | LBTESTCD | LBSTRESN | VISITNUM |
|---|---|---|---|---|
| YAMAA-01 | ALB | 4.0 | 1 | |
| YAMAA-01 | ALB | 4.5 | 2 | |
| YAMAA-01 | ALB | 5.5 | 4 | |
| YAMAA-01 | ALB | 7.0 | 5 | |
| YAMAA-01 | ALB | 4.0 | 1 | |
| YAMAA-01 | ALB | — | 2 | |
| YAMAA-01 | ALB | 5.0 | 3 | |
| YAMAA-01 | ALB | 3.75 | 1 | |
| YAMAA-01 | ALB | 4.25 | 1 | |
| YAMAA-01 | ALB | 4.75 | 2 | |
| YAMAA-01 | ALB | 4.0 | 1 | |
| YAMAA-01 | HGB | 13.5 | 1 | |
| YAMAA-01 | ALB | 4.5 | 1 | |
| YAMAA-01 | ALB | 4.5 | 2 | |
| YAMAA-01 | ALB | 5.0 | 3 |
Expected output
Under review13 expected rowsexpected/adlbc.csv
Edit13 rows| STUDYID | USUBJID | PARAMCD | AVISITN | AVAL | PREV_AVAL | CHG | PREV2 |
|---|---|---|---|---|---|---|---|
| YAMAA-01 | _ALB | 1 | 4 | — | — | — | |
| YAMAA-01 | _ALB | 2 | 4.5 | 4 | 0.5 | — | |
| YAMAA-01 | _ALB | 4 | 5.5 | 4.5 | 1 | 0.5 | |
| YAMAA-01 | _ALB | 5 | 7 | 5.5 | 1.5 | 1 | |
| YAMAA-01 | _ALB | 1 | 4 | — | — | — | |
| YAMAA-01 | _ALB | 3 | 5 | 4 | 1 | — | |
| YAMAA-01 | _ALB | 1 | 3.75 | — | — | — | |
| YAMAA-01 | _ALB | 1 | 4.25 | — | — | — | |
| YAMAA-01 | _ALB | 2 | 4.75 | 4.25 | 0.5 | — | |
| YAMAA-01 | _ALB | 1 | 4 | — | — | — | |
| YAMAA-01 | _ALB | 1 | 4.5 | — | — | — | |
| YAMAA-01 | _ALB | 2 | 4.5 | 4.5 | 0 | — | |
| YAMAA-01 | _ALB | 3 | 5 | 4.5 | 0.5 | 0 |
Comments and assessment materials
| Dataset Name | Variable Name | Variable Label | Type | Codelist/Controlled Terms | Core | Computational Method | Origin |
|---|---|---|---|---|---|---|---|
| ADLBC | STUDYID | Study Identifier | character | Copy value from LB.STUDYID. | Predecessor | ||
| ADLBC | USUBJID | Unique Subject Identifier | character | Copy value from LB.USUBJID. | Predecessor | ||
| ADLBC | PARAMCD | Parameter Code | character | Set per constructed row; see the Row construction sheet. | Derived | ||
| ADLBC | AVISITN | Analysis Visit Number | numeric | Copy value from LB.VISITNUM. | Predecessor | ||
| ADLBC | AVAL | Analysis Value | numeric | Copy value from LB.LBSTRESN. | Predecessor | ||
| ADLBC | PREV_AVAL | Previous Analysis Value | numeric | Take AVAL from 1 row earlier using window VISIT_ORDER. | Derived | ||
| ADLBC | CHG | Change from Previous Value | numeric | Compute AVAL - PREV_AVAL. | Derived | ||
| ADLBC | PREV2 | Lag of Change | numeric | Take CHG from 1 row earlier using window VISIT_ORDER. | Derived |
| Template | Include when | PARAMCD |
|---|---|---|
| change_alb | LB.LBTESTCD = 'ALB' AND LB.LBSTRESN IS NOT NULL | Constant "_ALB" |
| Version | Date | Author | Description | Reviewer | Sign-off |
|---|---|---|---|---|---|
| 1.3 | 2026-09-29 | yamaa docs build | Show study codelist names, external dictionaries, and ISO 8601 formats in variable rows; list the document's codelist values and metadata | ||
| 1.2 | 2026-09-29 | yamaa docs build | Classify ADaM copies from a declared input as Predecessor and recodes as Derived, and SDTM recodes of locally derived columns as Derived | ||
| 1.1 | 2026-09-29 | yamaa docs build | Resolve lookups into the donor record they select, and give every derivation shape a plain-language rule instead of its raw YAML | ||
| 1.0 | yamaa docs build | Initial generation from spec.yaml |
prompts/adam-adlbc-window-chain/full.md
Edit27 linesFollowing CDISC ADaM standards, use the provided LB dataset to create an ADLBC dataset with one record per subject per parameter per analysis visit.
The output dataset should contain the following columns in this order: STUDYID, USUBJID, PARAMCD, AVISITN, AVAL, PREV_AVAL, CHG, PREV2
Build the albumin change parameter "_ALB" from the albumin lab results; any other test produces no rows.
AVISITN is the analysis visit number, taken from the lab visit number. AVAL is the numeric albumin result for the visit.
PREV_AVAL is the value of the visit just before in visit order, for the same subject. It has no value on the subject's first visit.
CHG is the change since that visit, not the change from baseline; it has no value whenever PREV_AVAL has none. PREV2 is the CHG of the previous visit; it has no value until two earlier visits carry results.
A record with no numeric result gets no row, so the change and its lag compare against the latest earlier visit with a result. A zero change is a real value (0), distinct from no value.
Read the source datasets from /app/input and save the completed dataset as /app/output/adlbc.csv.
prompts/adam-adlbc-window-chain/conventions.md
Edit11 linesFollowing CDISC ADaM standards, use the provided LB dataset to create an ADLBC dataset with one record per subject per parameter per analysis visit.
The output dataset should contain the following columns in this order: STUDYID, USUBJID, PARAMCD, AVISITN, AVAL, PREV_AVAL, CHG, PREV2
PARAMCD is "_ALB" for albumin.
Read the source datasets from /app/input and save the completed dataset as /app/output/adlbc.csv.
prompts/adam-adlbc-window-chain/brief.md
Edit9 linesFollowing CDISC ADaM standards, use the provided LB dataset to create an ADLBC dataset with one record per subject per parameter per analysis visit.
The output dataset should contain the following columns in this order: STUDYID, USUBJID, PARAMCD, AVISITN, AVAL, PREV_AVAL, CHG, PREV2
Read the source datasets from /app/input and save the completed dataset as /app/output/adlbc.csv.
run.R
Edit4 lineslibrary(yamaa)adlbc <- yamaa_domain("spec.yaml")$outputadlbcrun.py
Edit4 linesimport yamaaadlbc = yamaa.yamaa_domain("spec.yaml").outputadlbcsolutions/adam-adlbc-window-chain/result.R
Edit40 lines# Reference solution for the yamaa benchmark adam-adlbc-window-chain (R track).## Written from the benchmark's full prompt and its inputs alone, as an agent# would write it. Harbor's oracle agent runs it to check that the prompt can# be solved and that the grader scores a correct answer 1.library(dplyr, warn.conflicts = FALSE)library(readr)lb <- read_csv( "/app/input/lb.csv", col_types = cols( LBSTRESN = col_double(), VISITNUM = col_integer(), .default = col_character() ))# Albumin results with a numeric value; a record without one gets no row,# so the lag compares against the latest earlier visit with a result.alb <- lb |> filter(LBTESTCD %in% "ALB", !is.na(LBSTRESN)) |> mutate(PARAMCD = "_ALB", AVISITN = VISITNUM, AVAL = LBSTRESN) |> arrange(STUDYID, USUBJID, AVISITN)# The change since the previous visit with a result, and its lag: the# previous visit's change. A zero change is a real value.adlbc <- alb |> group_by(STUDYID, USUBJID) |> mutate( PREV_AVAL = lag(AVAL), CHG = AVAL - PREV_AVAL, PREV2 = lag(CHG) ) |> ungroup() |> select(STUDYID, USUBJID, PARAMCD, AVISITN, AVAL, PREV_AVAL, CHG, PREV2) |> arrange(STUDYID, USUBJID, AVISITN)dir.create("/app/output", showWarnings = FALSE)write_csv(adlbc, "/app/output/adlbc.csv", na = "")solutions/adam-adlbc-window-chain/result.py
Edit43 lines# Reference solution for the yamaa benchmark adam-adlbc-window-chain (Python track).## Written from the benchmark's full prompt and its inputs alone, as an agent# would write it. Harbor's oracle agent runs it to check that the prompt can# be solved and that the grader scores a correct answer 1.from pathlib import Pathimport polars as pllb = pl.read_csv("/app/input/lb.csv", infer_schema=False).with_columns( pl.col("LBSTRESN").cast(pl.Float64), pl.col("VISITNUM").cast(pl.Int64),)# Albumin results with a numeric value; a record without one gets no row,# so the lag compares against the latest earlier visit with a result.alb = ( lb.filter(pl.col("LBTESTCD") == "ALB", pl.col("LBSTRESN").is_not_null()) .with_columns( PARAMCD=pl.lit("_ALB"), AVISITN=pl.col("VISITNUM"), AVAL=pl.col("LBSTRESN"), ) .sort(["STUDYID", "USUBJID", "AVISITN"]))# The change since the previous visit with a result, and its lag: the# previous visit's change. A zero change is a real value.adlbc = ( alb.with_columns( PREV_AVAL=pl.col("AVAL").shift(1).over(["STUDYID", "USUBJID"]), ) .with_columns(CHG=pl.col("AVAL") - pl.col("PREV_AVAL")) .with_columns(PREV2=pl.col("CHG").shift(1).over(["STUDYID", "USUBJID"])) .select( "STUDYID", "USUBJID", "PARAMCD", "AVISITN", "AVAL", "PREV_AVAL", "CHG", "PREV2", ))Path("/app/output").mkdir(exist_ok=True)adlbc.write_csv("/app/output/adlbc.csv")solutions/adam-adlbc-window-chain/result.sas
Edit28 lines/* opensas reference solution for the yamaa benchmark adam-adlbc-window-chain. Reads the benchmark inputs and derives the requested output. Execute with the existing opensas runtime: opensas /app/output/result.sas. */data lb; length STUDYID USUBJID LBTESTCD $1024; infile "/app/input/lb.csv" dsd dlm="," firstobs=2 truncover lrecl=32767; input STUDYID : $char1024. USUBJID : $char1024. LBTESTCD : $char1024. LBSTRESN VISITNUM;run;data collected; length PARAMCD $8; set lb; if LBTESTCD in ('ALB') and not missing(LBSTRESN); PARAMCD=cats('_',LBTESTCD); AVISITN=VISITNUM; AVAL=LBSTRESN;run;proc sort data=collected; by STUDYID USUBJID PARAMCD AVISITN; run;data result; retain previous priorchange; set collected; by STUDYID USUBJID PARAMCD; if first.PARAMCD then do; previous=.; priorchange=.; end; PREV_AVAL=previous; CHG=AVAL-PREV_AVAL; PREV2=priorchange; previous=AVAL; priorchange=CHG;run;proc sql; create table final as select STUDYID, USUBJID, PARAMCD, AVISITN, AVAL, PREV_AVAL, CHG, PREV2 from result;quit;proc export data=final outfile="/app/output/adlbc.csv" dbms=csv replace; run;
Comments
Signed in with GitHub; saved in repository Discussions