| --- |
| pretty_name: GymCL Datasets |
| license: other |
| tags: |
| - continual-learning |
| - agents |
| - tool-use |
| - code |
| - games |
| --- |
| |
| # GymCL 数据集(`Septzzz/gymcl-data`) |
|
|
| [GymCL](https://github.com/Septzzz/GymCL) 持续学习基准的数据。核心是 **`unified/`**: |
| 每个任务一份**统一格式的 golden `train` + `test`**,可直接用于 SFT / 离线评测。 |
|
|
| ## `unified/` — 统一 golden 数据(10 个任务) |
|
|
| 统一逐行 schema(train 与 test 同一种): |
| ```json |
| {"messages": [{"role":"...","content":"..."}], "reward": 1.0 或 null, "meta": {"task","env_id","instance","source", ...}} |
| ``` |
| - **train.jsonl** —— **golden 轨迹**:`messages` 含完整 `user/assistant(/tool)`,`reward=1.0`。 |
| - **test.jsonl** —— 输入侧 `messages`;有 gold 的把标准答案放 `meta.gold`,`meta.grader` 标明判分方式,`reward=null`。 |
|
|
| | 类 | 任务 | train golden 来源 | test | |
| |---|---|---|---| |
| | code | `swe_smith` | SWE-smith-trajectories(resolved 多轮轨迹,5000) | SWE-smith held-out(problem→patch gold) | |
| | code | `swebench_verified` | R2EGym-SFT-Trajectories(多轮,3231) | SWE-bench_Verified 500(problem→patch gold) | |
| | code | `terminal_bench` | Terminal-Bench **2.0** 轨迹(reward>0,5000) | TB2.0 任务(instruction→gold `solve.sh`,89) | |
| | tool | `toolbench_g1` | ToolLLaMA G123 DFS 轨迹(5000) | G1 query(LLM judge) | |
| | tool | `apibench_hf` | apibench huggingface_train(instruction→api_call gold,8191) | huggingface_eval(gold,911) | |
| | tool | `bfcl` | BFCL **16 个 gold 类别**(单轮 user→gold 调用 / 多轮逐轮调用,3200) | held-out 10%(gold in meta,356) | |
| | game | `mastermind` | **oracle**(一致性消元推理轨迹,200) | 50 eval seeds(env 评分) | |
| | game | `sokoban` | **oracle**(BFS 最优推箱,200) | 50 eval seeds | |
| | game | `fifteen_puzzle` | **oracle**(A* 最优解,200) | 50 eval seeds | |
| | game | `minesweeper` | **oracle**(首翻+标雷+清安全格,200) | 50 eval seeds | |
|
|
| 设计说明: |
| - **游戏无现成轨迹** → 用程序化 **oracle 求解器**在真环境里跑出最优 golden 轨迹(`env.step` 校验 reward)。 |
| - **bfcl 替换了 tau2**(tau2 无 golden 轨迹);**mbpp 已移除**(只有单步 gold、无多轮轨迹)。 |
| - **games / bfcl-test / toolbench-test** 由环境现场评分(引擎 reward / bfcl checker / LLM judge),故 `reward=null`、gold(若有)在 `meta`。 |
| - **terminal_bench 的实时评测**需 terminal-bench/harbor harness + Docker(不影响本数据的使用)。 |
| |
| ## 原生环境资产(供实际跑环境用,非 golden 数据) |
| |
| 部分环境运行时需要原生资产(工具 schema、实例 parquet 等),与 `unified/` 并列存放: |
| `SWE-smith/`、`toolbench/`、`apibench/` 等。运行时另有从公共 Hub 直接拉取的(R2E-Gym-Lite、MBPP、各轨迹池、TB2.0 任务集)。 |
| |
| ## 用法 |
| |
| ```bash |
| python scripts/download_data.py # GymCL setup.sh 会自动调用 |
| # 直接取统一 golden 数据做 SFT / 评测: |
| # unified/<task>/train.jsonl (golden messages, reward=1.0) |
| # unified/<task>/test.jsonl (输入 + meta.gold / 或待跑实例) |
| # unified/manifest.json (各任务 train/test 条数) |
| ``` |
| |