The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
model: string
reward: struct<detail: struct<carved_file: string>, pass: bool, reward: double>
child 0, detail: struct<carved_file: string>
child 0, carved_file: string
child 1, pass: bool
child 2, reward: double
seconds: double
stamp: string
steps: int64
stopped_because: string
task: string
verdict: string
verifier_rc: int64
oracle: struct<rc: int64, reward: double, tail: string>
child 0, rc: int64
child 1, reward: double
child 2, tail: string
empty: struct<rc: int64, reward: double, tail: string>
child 0, rc: int64
child 1, reward: double
child 2, tail: string
to
{'empty': {'rc': Value('int64'), 'reward': Value('float64'), 'tail': Value('string')}, 'oracle': {'rc': Value('int64'), 'reward': Value('float64'), 'tail': Value('string')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
model: string
reward: struct<detail: struct<carved_file: string>, pass: bool, reward: double>
child 0, detail: struct<carved_file: string>
child 0, carved_file: string
child 1, pass: bool
child 2, reward: double
seconds: double
stamp: string
steps: int64
stopped_because: string
task: string
verdict: string
verifier_rc: int64
oracle: struct<rc: int64, reward: double, tail: string>
child 0, rc: int64
child 1, reward: double
child 2, tail: string
empty: struct<rc: int64, reward: double, tail: string>
child 0, rc: int64
child 1, reward: double
child 2, tail: string
to
{'empty': {'rc': Value('int64'), 'reward': Value('float64'), 'tail': Value('string')}, 'oracle': {'rc': Value('int64'), 'reward': Value('float64'), 'tail': Value('string')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
swe-task-forge task pool
Every task, once. MANIFEST.jsonl has one forge-task.v1 envelope per task; tasks/<task_id>/
is a Harbor package (task.toml, instruction.md, environment/Dockerfile, tests/, solution/, runs/).
A task names its image by env_id; the bytes live in swe-task-forge-environments.
| tasks | attempted | solved | sources |
|---|---|---|---|
| 259 | 259 | 191 | build:fault_b2_20260906, build:fault_v1_20260905, carve:carve_b2_20260906, carve:carve_v1_20260905b, revert:mining_b2_20260906, swebench:mining_b2_20260906, swebench:recut_20260912, swebench:stream1_20260911, synth:recut2_20260912, synth:recut_20260912, synth:stream1_20260911 |
Benches are selections over this pool, by reference, in swe-task-forge-benches.
Batches
Every upload and every retirement, in order. A bench pins one pool_revision.
| when | action | campaign | pipeline tag | tasks | shapes | solved/attempted | pool revision | note |
|---|---|---|---|---|---|---|---|---|
| 2026-09-05 07:25 | upload | harbor_carve_20260905 | - |
59 | tests_present: 59 | 58/59 | 9cfa62d539aa |
v0: tests left in the container; glm-5.3 solved 58/59 (transcription, not solving) |
| 2026-09-05 08:05 | retire | harbor_carve_20260905 | - |
0 | 0/0 | 75715e8107a1 |
superseded by the withheld-tests shape | |
| 2026-09-05 08:21 | upload | carve_v1_20260905b | carve-v1.0.0 |
36 | withheld: 36 | 25/36 | fff8797e5dce |
|
| 2026-09-05 08:34 | upload | carve_v1_20260905b | carve-v1.0.0 |
36 | withheld: 36 | 26/36 | 36c85f820798 |
re-sync: 2 infra-failed attempts re-run on fixed code (1 solved, 1 step limit); v1 = 26/36 |
| 2026-09-05 14:04 | upload | fault_v1_20260905 | fault-v1.0.0 |
4 | fault: 4 | 3/4 | 6d644ceea01f |
fault line first tagged run: 139 candidates, 4 tasks, glm-5.3 solved 3; all from the build family |
| 2026-09-05 14:11 | upload | fault_v1_20260905 | fault-v1.0.0 |
4 | fault: 4 | 3/4 | bbcf19f8f65b |
fault line re-sync: tasks now carry route=build so benches can split the pool |
| 2026-09-05 16:46 | upload | carve_v1_20260905b | carve-v1.0.0 |
36 | withheld: 36 | 26/36 | 194b2f7fa603 |
instruction repair: the command block named a line lifted from the entry script; regenerated from each package own data |
| 2026-09-05 16:46 | upload | fault_v1_20260905 | fault-v1.0.0 |
4 | fault: 4 | 3/4 | 4935f63a4f83 |
instruction repair |
| 2026-09-07 15:33 | upload | fault_b2_20260906 | fault-v1.1.0 |
19 | fault: 19 | 19/19 | 2012a192ea82 |
batch-2: 55 images, 19 build faults, glm-5.3 19/19 solved -- the line is too easy, one shape only |
| 2026-09-07 18:50 | upload | carve_b2_20260906 | carve-v1.1.0 |
94 | withheld: 94 | 69/94 | 5400bc5888cc |
batch-2: 55 images, 94 carve tasks, glm-5.3 69/94 solved; .git sealed in the package (95 pct -> 73 pct) |
| 2026-09-07 18:59 | upload | mining_b2_20260906 | mining-v1.0.0 |
25 | suite_flip: 25 | 18/25 | 91e6c4c1f11c |
batch-2: 57 images with history, 25 commit fail-to-pass tasks, glm-5.3 18/25 solved; first batch this line has produced |
| 2026-09-07 20:04 | upload | fault_b2_20260906 | fault-v1.1.0 |
19 | fault: 19 | 19/19 | 5ca8bf13da5c |
re-upload: the first fault batch was packaged before the .git seal; these 19 replace them. 19/19 solved either way -- this shape hides its answer in /root/.forge-hidden/setup.sh, never in the history |
| 2026-09-13 07:32 | upload | minions_pack_20260913 | mining-v2.4.33-1-g0368ef2 |
81 | ?: 81 | 56/81 | 34af0cf82483 |
stream1_20260911 recut: 81 mining tasks over 6 images, audit 81/81, images digest-verified against the campaign |
数据说明(给要用这批数据的人)
这份文件在两个地方是同一份:仓库 docs/DATA_README.md,和 Hugging Face
Zeng-Weijun/swe-task-forge-tasks 的 README 下半部分。
每加一批数据就在最上面加一段:什么时候加的、哪个版本造的、为什么加、多少题、质量怎么样、smoke 结果。
上半部分的表格是上传器从盘上数据自动生成的,不用手改。
三个仓怎么分:-environments 放 docker 镜像(一个镜像只存一份),-tasks 放所有题("大池子"),-benches 只放 task_id 的选择列表。
一道题怎么用:docker load 它的镜像 → docker build -f environment/Dockerfile → agent 在里面干活(无网)→ 跑完把 tests/、solution/ 拷进去,root 跑 sh /tests/test.sh 看 reward.json。
2026-09-13 · minions(commit 挖矿线)· 81 题 · 6 个镜像(当前)
为什么加:这是"给一个 docker → 挖出题 → 真实 rollout → 环境问题回流"这条新链路(mining 一条链,
docs/PIPELINE.md 定形)跑出来的第一批可交付数据。输入是 repo2env 交付的 6 个 runtime 镜像,
campaign stream1_20260911(另有 5 题重切自 recut_20260912 / recut2_20260912)。
走的 pipeline:mining-v2.4.33。造完用 tools/audit_pack.py 过验收,81/81 通过:
每个包的 SHA256SUMS 覆盖到 environment/tree.tar(题目起始状态)和 GATE.json,
6 个基础镜像的 tarball 逐个和 campaign 真正跑的那个镜像按 digest 核验一致才打进来。
| 镜像 | 题 | 镜像 digest(短) |
|---|---|---|
| spesmilo/electrum | 52 | 0b462e7011ce |
| vitalik/django-ninja | 11 | 634a7adc34fa |
| Kaggle/kaggle-cli | 7 | b1e6b39365b5 |
| HIPS/autograd | 6 | 39783384527b |
| lmcinnes/umap | 3 | 66c85cec4f86 |
| hugapi/hug | 2 | 61e5e8e64aa7 |
| 合计 | 81 | codex+glm-5.3-flash 按会话解开 56/81 = 69% |
题面让 agent 做的是把被挖掉的那次提交重新做出来;判据是两臂门(空臂红、oracle 臂绿)。
每题都带我们自己跑的真实轨迹(runs/)。route 字段是 synth。
⚠️ 这批取代了 2026-09-08 那次没上传的
minions_pack_20260912——那份包images/是空的、tree.tar没进校验和,谁都构建不了。别用旧的。
怎么拉下来自己跑(简版;完整版见仓库 docs/PULL_FROM_HF.md)
数据集是公开的,不用 token。
pip install -U "huggingface_hub[cli]"
# 1. 拉清单,挑一个 task,记下它的 task_id 和 env_id
hf download Zeng-Weijun/swe-task-forge-tasks --repo-type dataset --include "MANIFEST.jsonl" --local-dir forge
TASK=synth.hips-autograd.mc4c9dcd05@39783384527b # 换成你挑的
ENV=$(python3 -c "import json;print(next(json.loads(l)['env_id'] for l in open('forge/MANIFEST.jsonl') if json.loads(l)['task_id']=='$TASK'))")
# 2. 拉 task 包 + 它的镜像
hf download Zeng-Weijun/swe-task-forge-tasks --repo-type dataset --include "tasks/$TASK/*" --local-dir forge
HF_HUB_ENABLE_HF_TRANSFER=1 hf download Zeng-Weijun/swe-task-forge-environments --repo-type dataset --include "$ENV/*" --local-dir envs
# 3. 载入镜像(大镜像切了 part,先拼回来)
cd "envs/$ENV"; [ -f image.tar.gz.part00 ] && cat image.tar.gz.part* > image.tar.gz; docker load < image.tar.gz; cd -
# 4. 构建 task 镜像,验两臂门:空臂必须红、打上标准答案必须绿
cd "forge/tasks/$TASK" && sha256sum -c SHA256SUMS
docker build -t forge-task -f environment/Dockerfile environment
docker run --rm --network none forge-task bash -lc 'bash /repo2env/test.sh'; echo "空臂 rc=$?(期望非 0)"
docker run --rm --network none forge-task bash -lc 'cd /workspace && (sh /solution/solve.sh || git apply -p1 /solution/gold.patch); bash /repo2env/test.sh'; echo "oracle rc=$?(期望 0)"
instruction.md 是题面,GATE.json 是两臂门证据,runs/ 是我们自己跑的真实轨迹。
2026-09-07 · batch-2 · 三条线一起 · 138 题(当前)
为什么加:v1 只有 15 个镜像。repo2env 交过来第二批仓库后有 55 个可用镜像,三条线在同一批镜像上各跑一遍, 第一次让 mining 线出了题。同时修掉一个会让整批数据作废的泄漏(下面第一条)。
走的 pipeline:lines/carve carve-v1.1.0、lines/fault fault-v1.1.0、lines/mining mining-v1.0.0,
三条线并行,campaign 分别是 carve_b2_20260906 / fault_b2_20260906 / mining_b2_20260906。
| 线 | 题 | 环境 | glm-5.3 解开 | 解开的中位步数 | 用满 30 步 |
|---|---|---|---|---|---|
| carve | 94 | 35 | 69 / 94 = 73% | 21 | 39 |
| fault | 19 | 19 | 19 / 19 = 100% | 13 | 0 |
| mining | 25 | 15 | 18 / 25 = 72% | 20 | 12 |
| 合计 | 138 | 40(去重) | 106 / 138 = 77% |
这批最重要的事:镜像里的 .git 被封掉了。
mining 线需要仓库历史,所以我们请 repo2env 在镜像里保留 .git(repo2env#71)。代价是 carve 的答案就在
git show HEAD:<被挖空的文件> 里——交付出去的 trace 里能直接看到模型这么做(git show HEAD:src/python.ts、
git show 8b06a60 --stat)。规则在申请 .git 的时候就写了,但一直没实现。现在打包时把 .git 移到
/root/.forge-sealed-git,容器里 git 命令直接 fatal: not a git repository(在真容器里验过)。
封之前 carve 的解开率是 95%,封之后是 73% —— 那 22 个百分点是模型在读答案。v1.0.0 及更早的 carve 数字按"历史可读"理解。
fault 线的包是第二次出的:第一次出包在封 .git 的提交之前,19 个包一个都没封(carve 94/94、mining 25/25 都封了,
只有 fault 漏了——它那条线跑在封漏之前,我一直没重新出包,并且已经先传了一次 HF)。
重新出包重跑:还是 19/19 全解开,中位 13 步。这条线的答案是被藏进 /root/.forge-hidden/setup.sh 的安装脚本,
本来就不在 git 历史里,所以封不封对它没影响——但池子里那 19 道已经被替换成封过的版本(账本里 replaced: 19 那一行)。
这些解开率量的是步数预算,不是题目难度。 carve 25 道没解开的里有 24 道(96%)用满了 30 步, mining 7 道里 5 道(71%);看 trace 它们不是在打转(命令重复率 0–27%),是跑到第 30 步还在读源码。 换句话说 "carve 73%" 应当读作 "30 步预算内 73%",把预算抬上去这个数就会动。三条线都是这样, 所有难度数字都带这个前提。另外 reward 只有 0.0/1.0 两个取值,没有部分得分。
质量与难度:
- carve 94 道里 39 道用满 30 步上限,解开的中位 21 步(v1 是 19 步)。难度形状对了:不是 bootstrap 死,是最后几个细节猜不到。
- mining 18/25,路数是
swebench20(commit 自带测试)+revert5。这是这条线第一次出题——v0.1.0 是 0 题。 (这批 mining 跑了两遍:第一遍的包直接抄了 carve 的task.toml,把 commit 标题写进carved_file,池子 id 里因此带上了空格和#。 修好 packager 后整条重出重跑,交付的是第二遍。第一遍是 20/25,第二遍 18/25——同一批题换一次掷骰子的差别。) - fault 19/19 全解开,是这批最弱的一环:19 道全是同一种形状(依赖被删 → 装回去),中位 12 步,没有区分度。 下一版要么换故障家族,要么这条线不该单独算一个难度档。
跑批时踩到的三个坑(都已修进代码,有回归测试):
relay failed after 4 attempts: 'choices'—— 94 道 carve 里 51 道是这个,全被记成"没解开", 于是第一版数字是 34/94 = 36%。根因是payload["choices"][0]抛KeyError,它的消息就是choices三个字, 中转站真正说了什么被丢掉了。现在报出中转站的原话,重试 6 次指数退避,遇到它一定会重复的拒绝就不再重试。- 补跑一个子集会把没重跑的那些结果删掉——
run_packages按这次跑了什么重写整个 manifest 和 tasks.jsonl。 现在 runner 会告诉 stage 输入被限制过,stage 做合并。 bundle_environments会把包移到environments/<env>/tasks/下,但上游 manifest 里的路径还是老的, 于是补跑时 51 道全报 "image does not build"。现在会去打包后的位置找。- 顺带修了:
tasks.jsonl里environment.digest一直是null(envelope 在问镜像 digest 之前就写好了), 138 道题全中。现在 bundling 拿到 digest 后重写 envelope。
已知问题:
- carve 掉得最多的一段还是
nothing_to_hollow177 / 424 —— 没有"该挖哪个模块"的判据,只是挨个试。 - mining 的
suite_indifferent_to_commit116 和gold_patch_does_not_apply76 仍是前两大掉落。 - fault 的 environment 家族 485/523 不咬。
2026-09-05 · fault 线 · fault-v1.0.0 · 4 题
为什么加:carve 线出的是"实现缺的模块",全是写代码的题。这一批是修复类:代码一个字没改,是机器坏了或依赖没了,agent 要找出原因修好。判官仍是仓库自己的套件。
走的 pipeline:lines/fault/。每个故障在镜像上先证三件事才算数——挖前套件绿、注入后红、跑修复脚本后又绿。两个家族:build(把依赖存放处删掉,同时把 /repo2env/setup.sh 藏起来,因为它就是答案;镜像里的离线缓存留着)和 environment(改 locale/时区/加载器路径/名字解析这些,代码不动)。
多少:15 个 docker × 各语言可用故障 = 139 个候选 → 5 个真咬 → 4 个过两臂门。
质量(smoke,glm-5.3):3 解开 / 1 没解开。全部 4 道来自 build 家族。
已知问题:environment 家族 128/134 不咬——这批仓库的套件只跑纯逻辑,不碰 locale、时区、加载器、名字解析。这条线的产量取决于"套件碰多少机器",不是故障库有多大;要提产量得按证据设计故障(先问容器里有什么、看套件运行时真正加载了哪些库),而不是按目录穷举。
2026-09-05 · v1 · carve-v1.0.0 · 36 题(当前)
为什么加:第一批(v0)把仓库自带的测试留在了容器里,模型直接对着测试抄,98% 全过,量不出能力。这一批把覆盖该模块的测试跟代码一起抽走,题面明说"没有测试,写测试是任务的一部分",判分用抽走的原版测试。
走的 pipeline:carve(挖空一个文件的所有函数体,仓库自己的套件当判官)。六段:列候选 → 抽测试 → 打 Harbor 包 → 两臂门(不给答案必挂、给 oracle 必过)→ 真 agent 跑一遍 → 按环境打包。代码 tag carve-v1.0.0。
多少:15 个 docker 进,68 个候选 → 43 个找得到覆盖测试 → 36 个过两臂门 → 10 个环境交付。7 个仓库有产出,产量最多的是 archtechx/enums(PHP)和 darkrishabh/agent-skills-eval(TypeScript)。
质量(smoke,glm-5.3,并发 12,无网):26 解开 / 10 没解开 = 72%,中位 19 步。(首轮是 25/9/2,2 道基础设施失败用修好的代码重试后一道解开一道到 30 步上限,已并入。)
对照 v0 的 98% / 13 步。没解开的题看 trace 都是同一种情况:实现了、自己写的测试过了、现有套件也绿,但原作者测试盯的细节没猜到(比如异常消息的确切措辞、边界值)。这是想要的难度形状——不是 bootstrap 直接死。(说明一句:reward 只有 0.0 和 1.0 两个取值,所以"部分可得分"是我们看 trace 得到的印象,不是 reward 量出来的。)
一个好例子:annamedv/red 的 levenshtein.go,agent 从调用方反推语义,自己写了 4 个测试(含参考实现交叉验证和边界测试),go vet + -race,12 步过隐藏测试。
已知问题:
- 首轮 2 道"基础设施问题"是我们自己的 bug(agent 清了
/tmp让判官脚本落不了地;一道题反复跑要联网的npm test拖了 45 分钟)。都已修进carve-v1.0.0,重试结果已并入上面的数字(池子 2026-09-05 第二次同步,账本有一行 note)。 - 6 个候选在两臂门掉在"不给答案也过"——找到的"覆盖测试"其实没覆盖到,门把它们拦了,没进池子。
- 没有 C 语言的挖空后端,两个 C 仓库零产出。
2026-09-05 · v0 · 59 题(已下线)
第一批,测试留在容器里。glm-5.3 解开 58/59、中位 13 步;读 60 个 trace,23 个是先 cat 完整测试再写代码,其余靠套件断言报错反推——是抄规格不是解题。所以下线了:池子里的 59 个 task 目录已删,镜像保留给 v1 复用。留着这一段是为了说明 v1 为什么长那样。
- Downloads last month
- 622