| # lg-bench-general-v1.0: 35 general tasks |
|
|
| English | [中文](README_zh.md) |
|
|
| General enterprise office tasks. Difficulty: L1 9, L2 9, L3 9, L4 8. |
|
|
| | Prefix | Tasks | Scope | |
| | --- | ---: | --- | |
| | `CG` | 2 | Multi-file merging, batch template filling | |
| | `CPV` | 3 | Expense compliance, policy conflicts, requirement feasibility checks | |
| | `DA` | 5 | Business data analysis, attribution, and rolling forecasts | |
| | `DFP` | 4 | Data cleaning, customer segmentation, Text2SQL, financial intelligence screening | |
| | `DIAG` | 4 | Error logs, root cause, and integration failure diagnosis | |
| | `DRV` | 5 | Contract clauses, compliance red lines, form and multi-invoice verification | |
| | `IMG` | 2 | Invoice and contract image recognition and reconciliation | |
| | `KBR` | 4 | Multi-source retrieval over the Kingdee Cloud Community knowledge base | |
| | `RQA` | 3 | Q&A over annual reports and ESG reports | |
| | `SCH` | 3 | Calendar sync, time zone conversion, date rules | |
|
|
| Each task's inputs live in `environment/` (some tasks pack them as `workspace.tar.gz`), gold answers and judge evidence live in `tests/gold/`, and the source materials are recorded in `source_dataset` / `source_item` of `task.toml`. |
|
|
| ## Scoring |
|
|
| Each task's `tests/judge.yaml` combines rule-based sub-scores with an LLM judge (`rubric_binary_mean`). The judge reads `JUDGE_BASE_URL`, `JUDGE_MODEL`, and `JUDGE_API_KEY` (OpenAI-compatible API) and runs in strict mode by default: when the judge is unavailable, the result must be reported separately, and must not be counted as a business score of zero or replaced with the rule-based score. |
|
|
| In official runs, task containers are attached only to an internal network; only the model proxy has external access. |
|
|
| ## Integrity |
|
|
| [`dataset.toml`](dataset.toml) records each task's name and Harbor task package digest. |
|
|
