---
configs:
- config_name: rollouts
  data_files: [{split: train, path: "rollouts/**/*.parquet"}]
- config_name: executions
  data_files: [{split: train, path: "executions/**/*.parquet"}]
- config_name: exec_outputs
  data_files: [{split: train, path: "exec_outputs/**/*.parquet"}]
- config_name: exec_inputs
  data_files: [{split: train, path: "exec_inputs/**/*.parquet"}]
- config_name: problems
  data_files: [{split: train, path: "problems/**/*.parquet"}]
---

# mlcb-rollouts — multilingual LiveCodeBench (Agnostics stdin/stdout)

Bare-model Qwen3 4B/8B/14B (thinking, t=0.6) rollouts on stdin/stdout LiveCodeBench for fortran and cobol.
Tables (config dropdown): rollouts (generations + passed), executions (per-test verdicts/error_code),
exec_outputs (dedup stdout), exec_inputs (dedup stdin), problems. language/model/temp are columns in
rollouts+executions; filter in the viewer SQL console.