| --- |
| license: apache-2.0 |
| language: [en] |
| pretty_name: terminal-bench-mini |
| tags: [benchmark, coding-agents, terminal-bench, subset, evaluation] |
| size_categories: [n<1K] |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: data/tasks.jsonl |
| --- |
| |
| # terminal-bench-mini |
|
|
| Fourteen of Terminal-Bench 2.0's ninety tasks, picked so that ranking agents on |
| the subset reproduces ranking them on the whole benchmark. |
|
|
| Running ninety tasks five times each is how the official leaderboard is built. |
| That is out of reach if you are comparing quant variants, fine-tunes or local |
| models on your own hardware. This subset turns a multi-day sweep into a few |
| hours. |
|
|
| Same approach as [deepswe-mini](https://huggingface.co/datasets/LocalLLaMA/deepswe-mini): |
| take the published per-task results, rank the field on the full benchmark, then |
| find a small subset that preserves that ranking. **These fourteen tasks |
| reproduce about 87% of the full benchmark's pairwise ordering, and 97% of it |
| when two agents are more than ten points apart.** |
|
|
| ## The tasks |
|
|
| | task | field pass rate | spread across agents | run-to-run noise | |
| |---|---|---|---| |
| | filter-js-from-html | 11.9% | 0.310 | 0.022 | |
| | install-windows-3.11 | 11.9% | 0.258 | 0.090 | |
| | mteb-retrieve | 41.2% | 0.420 | 0.149 | |
| | db-wal-recovery | 56.8% | 0.448 | 0.097 | |
| | qemu-alpine-ssh | 69.3% | 0.417 | 0.093 | |
| | build-pmars | 75.2% | 0.361 | 0.122 | |
| | mcmc-sampling-stan | 81.5% | 0.328 | 0.100 | |
| | rstan-to-pystan | 81.7% | 0.329 | 0.102 | |
| | kv-store-grpc | 86.4% | 0.297 | 0.063 | |
| | bn-fit-modify | 87.6% | 0.295 | 0.056 | |
| | regex-log | 90.8% | 0.270 | 0.025 | |
| | custom-memory-heap-crash | 91.9% | 0.273 | 0.000 | |
| | crack-7z-hash | 93.5% | 0.228 | 0.022 | |
| | vulnerable-secret | 94.6% | 0.226 | 0.000 | |
|
|
| Field pass rate is the mean over 37 agents. Spread is the standard deviation of |
| those per-agent rates: a task with spread near zero tells you nothing about who |
| is better. Run-to-run noise is the average standard deviation across repeated |
| trials of the same agent on the same task. |
|
|
| Task definitions live in Terminal-Bench 2.0. This dataset names them, it does |
| not redistribute them. |
|
|
| ## Method |
|
|
| The source is every trial published to the Terminal-Bench 2.0 leaderboard: |
| **32,803 trials, 27,405 of them scorable**, across 90 tasks and 75 agent/model |
| submissions. Trials whose environment failed to start carry `exception_info` and |
| no verifier result; they are dropped rather than counted as failures. The 37 |
| submissions that attempted at least 85 tasks form the reference ranking. |
|
|
| The subset is a stratified sample. Tasks that every agent passes or every agent |
| fails carry no ranking information and are excluded first. The rest are sorted |
| by field pass rate into seven difficulty bands; within each band the quieter |
| half is preferred, since a task that flips between identical runs costs every |
| future user extra trials to average out; two tasks are drawn per band with a |
| fixed seed (20260919). |
|
|
| Quality is measured as pairwise agreement: over every pair of agents, how often |
| the subset orders them the way all ninety tasks do, broken out by how far apart |
| the pair really is. |
|
|
| | pair separation | agreement | |
| |---|---| |
| | more than 10 points | 97.3% | |
| | 5 to 10 points | 80.1% | |
| | 2 to 5 points | 73.2% | |
| | under 2 points | 65.1% | |
|
|
| Agreement also rises with how many tasks you run, which `analysis/size_curve.csv` |
| records: 14 tasks average 85%, 22 average 89%, 30 average 91%, 60 average 95%. |
| If the systems you are comparing are close, add tasks or add trials. |
|
|
| ## What it is good for |
|
|
| Use it to answer "is this configuration roughly competitive". The top of the |
| table is stable — the leading agent leads on both, and every agent in the full |
| top ten stays in the mini top twelve. Maximum rank movement is 16 places, and it |
| happens in the crowded middle where full-benchmark scores differ by fractions of |
| a point. |
|
|
| Do not use it to claim one agent beats another by two points. Five of the |
| fourteen tasks have run-to-run noise above 0.09, so run more than one trial per |
| task or expect a few points of movement that mean nothing. |
|
|
| ## Files |
|
|
| - `data/tasks.jsonl` — the fourteen tasks with their field statistics |
| - `analysis/leaderboard.csv` — all 37 agents scored on the mini set and the full 90, with both ranks and the movement between them |
| - `analysis/mini_trials.csv` — the agent × task pass rates behind that leaderboard |
| - `analysis/size_curve.csv` — agreement against subset size, 200 random subsets per size |
|
|
| ## Reproducing |
|
|
| Source data is the public `harborframework/terminal-bench-2-leaderboard` |
| repository. Every trial record is a small `result.json` sitting next to terminal |
| logs that make the repository 121 GB, so fetch only the records: |
|
|
| ```bash |
| git clone --filter=blob:none --no-checkout \ |
| https://huggingface.co/datasets/harborframework/terminal-bench-2-leaderboard |
| cd terminal-bench-2-leaderboard |
| git sparse-checkout set --no-cone '/**/result.json' |
| git checkout |
| ``` |
|
|
| That is 348 MB and about two minutes. Each record carries |
| `verifier_result.rewards.reward`. Note that two task-name conventions coexist in |
| the repository, `terminal-bench/<name>` and `<name>`; normalise them or the same |
| task counts twice and no two agents ever overlap. |
|
|
| ## Caveats |
|
|
| Agreement is measured against Terminal-Bench 2.0 as scored by these 37 |
| submissions, all of them frontier hosted models. A small local model sits below |
| that range, where the subset is untested — though a gap that large is the case |
| it handles most reliably. |
|
|
