|
Download tasks/task_021/instruction.md from OpenMOSS-Team/SWE-bench-Science: direct link, hf CLI and curl.
- Browser
- Download file 1.32 kB
-
https://huggingface.co/datasets/OpenMOSS-Team/SWE-bench-Science/resolve/main/tasks/task_021/instruction.md
- Command line
-
hf download hf://datasets/OpenMOSS-Team/SWE-bench-Science/tasks/task_021/instruction.md
-
curl -L -o instruction.md https://huggingface.co/datasets/OpenMOSS-Team/SWE-bench-Science/resolve/main/tasks/task_021/instruction.md
1.32 kB
| # Complete a Missing Atomic-Liquid Thermodynamic-Integration Capability | |
| The supplied source is a snapshot of a real research workflow for atomic-liquid | |
| Hamiltonian thermodynamic integration. It generates simulation inputs for a | |
| coupling path between a reference interaction and a target atomistic model. | |
| Some scientifically valid workloads are not represented consistently by the | |
| current implementation. The discrepancy is easy to miss in an ordinary | |
| interior case, but becomes visible when the coupling path reaches its boundary | |
| or when the system contains more than one chemical element with non-identical | |
| interaction parameters. | |
| Use the source snapshot and public generated-input artifacts to investigate the | |
| scientific inconsistency and complete the missing | |
| workflow capability. Preserve the existing public APIs and behavior outside the | |
| affected scientific cases. The implementation must generalize across supported | |
| element counts and parameter representations; do not hard-code the public | |
| fixtures or their numeric values. | |
| The public reproduction is generation-only and offline. It does not require a | |
| LAMMPS executable, a machine-learning model, a scheduler, network access, or | |
| external files. The benchmark evaluates the scientific behavior of the generated | |
| workflow rather than a particular textual patch. | |