Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
21
64
repository
stringlengths
8
51
repository_url
stringlengths
27
70
base_commit
stringlengths
40
40
upstream_commit
stringlengths
40
40
request
stringlengths
40
22k
size
stringclasses
3 values
changed_lines
int64
10
389
language
stringclasses
1 value
source_files
listlengths
1
6
licence
stringclasses
4 values
licence_file
stringlengths
35
78
licence_source_file
stringclasses
4 values
licence_sha256
stringlengths
64
64
other_licences
listlengths
0
6
spec
stringlengths
6.05k
25.9k
spec_path
stringlengths
35
78
spec_rewritten
bool
2 classes
spec_tests_file
stringlengths
16
50
spec_tests_pass_with_upstream_change
bool
2 classes
upstream_test_files
listlengths
1
10
hidden_tests
listlengths
1
10
changed_tests
listlengths
1
40
unchanged_tests
listlengths
0
376
has_trained_patch
bool
2 classes
trained_patch
stringlengths
2.14k
26.4k
⌀
trained_patch_path
stringlengths
41
84
⌀
trained_patch_model
stringclasses
1 value
trained_result
dict
snapshot
dict
Kludex__starlette__1766228b4e
Kludex/starlette
https://github.com/Kludex/starlette
a579222b161e5407cd311c32648ed7c845bda4a4
1766228b4eb5462efd552f3364656e1e0a5b0180
Match host names case-insensitively in `Host` routing and `TrustedHostMiddleware` (#3542)
medium
10
python
[ "starlette/middleware/trustedhost.py", "starlette/routing.py" ]
BSD-3-Clause
tasks/Kludex__starlette__1766228b4e/LICENSE
LICENSE.md
dcb95677a02240243187e964f941847d19b17821cf99e5afae684fab328c19bf
[]
# Build spec: case-insensitive host matching in `Host` routing and `TrustedHostMiddleware` ## Context `starlette/routing.py` defines `compile_path()`, which builds the regex used by both path routes and host routes. For host strings it strips the port and returns a regex compiled with no flags, so matching is case-se...
tasks/Kludex__starlette__1766228b4e/spec.md
true
tests/test_routing.py
true
[ "tests/middleware/test_trusted_host.py", "tests/test_routing.py" ]
[ { "path": "tasks/Kludex__starlette__1766228b4e/hidden/tests/middleware/test_trusted_host.py", "repo_path": "tests/middleware/test_trusted_host.py", "bytes": 3551, "sha256": "8086904f4094cc3a38d91fa8b7f55f965446f1d1cd85911954eb07cbb0bea8dc", "content": "from __future__ import annotations\n\nimpor...
[ "tests/middleware/test_trusted_host.py::test_trusted_host_middleware_is_case_insensitive", "tests/test_routing.py::test_host_routing_is_case_insensitive" ]
[ "tests/middleware/test_trusted_host.py::test_trusted_host_middleware", "tests/middleware/test_trusted_host.py::test_trusted_host_middleware_ipv6", "tests/middleware/test_trusted_host.py::test_default_allowed_hosts", "tests/middleware/test_trusted_host.py::test_www_redirect", "tests/test_routing.py::test_rou...
true
--- a/starlette/middleware/trustedhost.py +++ b/starlette/middleware/trustedhost.py @@ -25,7 +25,7 @@ if pattern.startswith("*") and pattern != "*": assert pattern.startswith("*."), ENFORCE_DOMAIN_WILDCARD self.app = app - self.allowed_hosts = list(allowed_hosts) + s...
tasks/Kludex__starlette__1766228b4e/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{ "seed": 1, "passed_spec_tests": true, "spec_tests": 9, "passed_upstream_tests": false, "tool_calls": 23, "wall_seconds": 195.3, "timed_out": false, "turn_limit": 30, "time_limit_seconds": 480, "generated_files_dropped": 0 }
{ "path": "tasks/Kludex__starlette__1766228b4e/repo.tar.gz", "files": 151, "bytes": 4050552, "compressed_bytes": 2733066, "sha256": "4616d445f0d6d400c7033c795d7a46fe1addf2402ea7138ef6156387888dcfeb", "omitted": [] }
Kludex__starlette__6beaa2e39d
Kludex/starlette
https://github.com/Kludex/starlette
ee93d2960b52c4e7c4a7abd40eb50773f7fad01e
6beaa2e39d3381e1d7ddb89d4cc7cbb8c7ac0961
Preserve response trailers through GZipMiddleware (#3565)
small
10
python
[ "starlette/middleware/gzip.py" ]
BSD-3-Clause
tasks/Kludex__starlette__6beaa2e39d/LICENSE
LICENSE.md
dcb95677a02240243187e964f941847d19b17821cf99e5afae684fab328c19bf
[]
# Build spec: Preserve response trailers through GZipMiddleware ## Context `starlette/middleware/gzip.py` implements `GZipMiddleware`, which wraps an ASGI app in either `GZipResponder` (when the client sends `Accept-Encoding: gzip`) or `IdentityResponder`, and rewrites outgoing `http.response.start` / `http.response....
tasks/Kludex__starlette__6beaa2e39d/spec.md
false
tests/middleware/test_gzip.py
true
[ "tests/middleware/test_gzip.py" ]
[ { "path": "tasks/Kludex__starlette__6beaa2e39d/hidden/tests/middleware/test_gzip.py", "repo_path": "tests/middleware/test_gzip.py", "bytes": 19463, "sha256": "9baa8000fe629359e4e6213079f7155e139841f1fd63427aa6d0f719e28eca3f", "content": "from __future__ import annotations\n\nimport hashlib\nimpo...
[ "tests/middleware/test_gzip.py::test_gzip_cleared_exclude_content_types", "tests/middleware/test_gzip.py::test_gzip_trailers", "tests/middleware/test_gzip.py::test_gzip_excluded_content_type_trailers" ]
[ "tests/middleware/test_gzip.py::test_gzip_responses", "tests/middleware/test_gzip.py::test_gzip_not_in_accept_encoding", "tests/middleware/test_gzip.py::test_gzip_ignored_for_small_responses", "tests/middleware/test_gzip.py::test_gzip_streaming_response", "tests/middleware/test_gzip.py::test_gzip_compressio...
false
null
null
null
null
{ "path": "tasks/Kludex__starlette__6beaa2e39d/repo.tar.gz", "files": 150, "bytes": 4020466, "compressed_bytes": 2726996, "sha256": "5babb31c381b2b835d1da781351fb0589a363f7ccfdd4d65c00a74241fde5970", "omitted": [] }
anthropics__anthropic-sdk-python__55d54fb2b3
anthropics/anthropic-sdk-python
https://github.com/anthropics/anthropic-sdk-python
1394cfcdc65cf14a5fba2c9e768760b4a893da46
55d54fb2b3e7d7308e136d987892fdf0a4278827
fix(client): allow API errors to be unpickled and copied Stainless-Generated-From: 1d39a96381717940589ec65f32437229e486b3e9
small
12
python
[ "src/anthropic/_exceptions.py" ]
Apache-2.0
tasks/anthropics__anthropic-sdk-python__55d54fb2b3/LICENSE
LICENSE
7fc343be7ec189b803c1872267cbaf0381b12ffdb22d972b8ea7b21b262ba563
[ { "path": "LICENSE-MIT", "licence": "Apache-2.0 AND MIT" }, { "path": "src/anthropic/_vendor/httpx_aiohttp/LICENSE", "licence": "BSD-3-Clause" } ]
## Context `src/anthropic/_exceptions.py` defines the SDK's exception hierarchy. Every error the client raises derives from `AnthropicError`: ```python class AnthropicError(Exception): pass ``` The API error classes take required constructor arguments, so the default exception pickling, which calls `cls(*self.ar...
tasks/anthropics__anthropic-sdk-python__55d54fb2b3/spec.md
false
tests/test_exceptions.py
true
[ "tests/test_exceptions.py" ]
[ { "path": "tasks/anthropics__anthropic-sdk-python__55d54fb2b3/hidden/tests/test_exceptions.py", "repo_path": "tests/test_exceptions.py", "bytes": 3613, "sha256": "514322314952637d3a56c731d8e45af2fe4c2895b59780899d77cda7fd594b01", "content": "from __future__ import annotations\n\nimport copy\nimp...
[ "tests/test_exceptions.py::test_duplicate_has_the_class_and_message_of_the_error", "tests/test_exceptions.py::test_duplicate_of_a_raised_status_error_keeps_its_attributes", "tests/test_exceptions.py::test_pickled_status_error_loads_as_its_own_class", "tests/test_exceptions.py::test_error_raised_in_a_process_p...
[]
true
--- a/src/anthropic/_exceptions.py +++ b/src/anthropic/_exceptions.py @@ -39,8 +39,19 @@ ] +def _reconstruct_error( + cls: type[AnthropicError], args: tuple[object, ...], state: dict[str, Any] +) -> AnthropicError: + """Rebuild an exception of `cls` without calling its `__init__`.""" + error = cls.__new__...
tasks/anthropics__anthropic-sdk-python__55d54fb2b3/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{ "seed": 3, "passed_spec_tests": true, "spec_tests": 7, "passed_upstream_tests": true, "tool_calls": 16, "wall_seconds": 356.3, "timed_out": false, "turn_limit": 30, "time_limit_seconds": 480, "generated_files_dropped": 0 }
{ "path": "tasks/anthropics__anthropic-sdk-python__55d54fb2b3/repo.tar.gz", "files": 3203, "bytes": 14354347, "compressed_bytes": 2376672, "sha256": "260adea078a7775f681dd0b32a0e418c3436928117f27692974dbf1a074842ae", "omitted": [] }
anthropics__anthropic-sdk-python__b0a7c23285
anthropics/anthropic-sdk-python
https://github.com/anthropics/anthropic-sdk-python
34faa4f12b8813da58925696648ccdfd403a7048
b0a7c23285972077462d889ed8f75ec424b117c2
fix(parse): keep const in structured output schemas, and drop type beside const and enum (#976)
small
26
python
[ "src/anthropic/lib/_parse/_transform.py" ]
Apache-2.0
tasks/anthropics__anthropic-sdk-python__b0a7c23285/LICENSE
LICENSE
7fc343be7ec189b803c1872267cbaf0381b12ffdb22d972b8ea7b21b262ba563
[ { "path": "LICENSE-MIT", "licence": "Apache-2.0 AND MIT" }, { "path": "src/anthropic/_vendor/httpx_aiohttp/LICENSE", "licence": "BSD-3-Clause" } ]
# Build spec: keep `const` in structured output schemas, drop `type` beside `const` and `enum` ## Context `src/anthropic/lib/_parse/_transform.py` turns a pydantic model or a dict JSON schema into the strict schema the Messages API expects. The change touches only `transform_schema` in that file; it is called from `t...
tasks/anthropics__anthropic-sdk-python__b0a7c23285/spec.md
true
tests/lib/parse/test_transform.py
true
[ "tests/lib/_parse/test_transform.py" ]
[ { "path": "tasks/anthropics__anthropic-sdk-python__b0a7c23285/hidden/tests/lib/_parse/test_transform.py", "repo_path": "tests/lib/_parse/test_transform.py", "bytes": 7760, "sha256": "2209e92a574cfce25dc93d328877006821aa10a38ac049c911ec4e3ffcd37c92", "content": "from copy import deepcopy\nfrom ty...
[ "tests/lib/_parse/test_transform.py::test_ref_schema_with_defs", "tests/lib/_parse/test_transform.py::test_enum_schema", "tests/lib/_parse/test_transform.py::test_keeps_const_and_enum_and_drops_type", "tests/lib/_parse/test_transform.py::test_moves_unsupported_const_and_enum_into_description", "tests/lib/_p...
[ "tests/lib/_parse/test_transform.py::test_ref_schema", "tests/lib/_parse/test_transform.py::test_anyof_schema", "tests/lib/_parse/test_transform.py::test_allof", "tests/lib/_parse/test_transform.py::test_object_schema", "tests/lib/_parse/test_transform.py::test_array_schema", "tests/lib/_parse/test_transf...
true
--- a/src/anthropic/lib/_parse/_transform.py +++ b/src/anthropic/lib/_parse/_transform.py @@ -144,6 +144,9 @@ any_of = json_schema.pop("anyOf", None) one_of = json_schema.pop("oneOf", None) all_of = json_schema.pop("allOf", None) + enum = json_schema.pop("enum", None) + has_const = "const" in json_...
tasks/anthropics__anthropic-sdk-python__b0a7c23285/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{ "seed": 2, "passed_spec_tests": true, "spec_tests": 11, "passed_upstream_tests": false, "tool_calls": 16, "wall_seconds": 221.3, "timed_out": false, "turn_limit": 30, "time_limit_seconds": 480, "generated_files_dropped": 0 }
{ "path": "tasks/anthropics__anthropic-sdk-python__b0a7c23285/repo.tar.gz", "files": 3210, "bytes": 14360976, "compressed_bytes": 2378564, "sha256": "c925bf0e9a840ca66b6ffd604ca7dcf020079720e22acab418d6816b134cf5e6", "omitted": [] }
ekalinin__nodeenv__b68eb29e6f
ekalinin/nodeenv
https://github.com/ekalinin/nodeenv
05e06b83e743cb772a8ec864775d7d175a812656
b68eb29e6f01b05c7920bd690047d8ca038c4701
"fix(nodeenv): do not put the venv bin on PATH twice with -p (#448)\n\nThe python venv's own activat(...TRUNCATED)
small
29
python
[ "nodeenv.py" ]
BSD-2-Clause
tasks/ekalinin__nodeenv__b68eb29e6f/LICENSE
LICENSE
606faf42d48b54d539dae99db6fecd48544535587e27275b319a3f36113cb6fa
[]
"## Context\n\n`nodeenv.py` is the whole package: it creates node.js environments and writes activat(...TRUNCATED)
tasks/ekalinin__nodeenv__b68eb29e6f/spec.md
true
tests/test_activate_path_dedup.py
false
[ "tests/test_activate_shells.py", "tests/test_install_activate.py" ]
[{"path":"tasks/ekalinin__nodeenv__b68eb29e6f/hidden/tests/test_activate_shells.py","repo_path":"tes(...TRUNCATED)
["tests/test_activate_shells.py::test_python_virtualenv_puts_bin_on_path_once","tests/test_install_a(...TRUNCATED)
["tests/test_activate_shells.py::test_syntax","tests/test_activate_shells.py::test_activate_sets_env(...TRUNCATED)
false
null
null
null
null
{"path":"tasks/ekalinin__nodeenv__b68eb29e6f/repo.tar.gz","files":37,"bytes":496208,"compressed_byte(...TRUNCATED)
eliben__pycparser__6a4459247a
eliben/pycparser
https://github.com/eliben/pycparser
7217a09643f32ebda208b03968b9039248db5f5b
6a4459247a2f35ea55b44028250d1bed6d2dfac4
Fix parser for declarator-less members Closes #633
small
13
python
[ "pycparser/c_parser.py" ]
BSD-3-Clause
tasks/eliben__pycparser__6a4459247a/LICENSE
LICENSE
0a9e78156ae8b96caf1b4f11ac2df424b83de29736f9e757630548a6272fdf90
[]
"# Build spec: fix parsing of declarator-less members\n\n## Context\n\n`pycparser/c_parser.py` is th(...TRUNCATED)
tasks/eliben__pycparser__6a4459247a/spec.md
true
tests/test_declarator_less_members.py
false
[ "tests/test_c_parser.py" ]
[{"path":"tasks/eliben__pycparser__6a4459247a/hidden/tests/test_c_parser.py","repo_path":"tests/test(...TRUNCATED)
[ "tests/test_c_parser.py::TestCParser_fundamentals::test_struct_member_without_declarator" ]
["tests/test_c_parser.py::TestCParser_fundamentals::test_FileAST","tests/test_c_parser.py::TestCPars(...TRUNCATED)
false
null
null
null
null
{"path":"tasks/eliben__pycparser__6a4459247a/repo.tar.gz","files":211,"bytes":1388942,"compressed_by(...TRUNCATED)
eliben__pycparser__712ba101f2
eliben/pycparser
https://github.com/eliben/pycparser
e174787fe31f9f342747c79f22d5f742306dce61
712ba101f2c860090d8719352fed9b4bf9b536a3
Fix generation for atomic members w/o declarators
small
24
python
[ "pycparser/c_generator.py" ]
BSD-3-Clause
tasks/eliben__pycparser__712ba101f2/LICENSE
LICENSE
0a9e78156ae8b96caf1b4f11ac2df424b83de29736f9e757630548a6272fdf90
[]
"## Context\n\n`pycparser/c_generator.py` holds `CGenerator`, the AST-to-C visitor. `pycparser/c_par(...TRUNCATED)
tasks/eliben__pycparser__712ba101f2/spec.md
true
tests/test_c_generator_atomic_members.py
true
[ "tests/test_c_generator.py" ]
[{"path":"tasks/eliben__pycparser__712ba101f2/hidden/tests/test_c_generator.py","repo_path":"tests/t(...TRUNCATED)
[ "tests/test_c_generator.py::TestCtoC::test_atomic_members_without_declarators" ]
["tests/test_c_generator.py::TestFunctionDeclGeneration::test_partial_funcdecl_generation","tests/te(...TRUNCATED)
true
"--- a/pycparser/c_generator.py\n+++ b/pycparser/c_generator.py\n@@ -699,6 +699,29 @@\n (...TRUNCATED)
tasks/eliben__pycparser__712ba101f2/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{"seed":2,"passed_spec_tests":true,"spec_tests":8,"passed_upstream_tests":false,"tool_calls":8,"wall(...TRUNCATED)
{"path":"tasks/eliben__pycparser__712ba101f2/repo.tar.gz","files":211,"bytes":1392215,"compressed_by(...TRUNCATED)
eliben__pycparser__878c289076
eliben/pycparser
https://github.com/eliben/pycparser
840fe55e46df0d8ba956057a286526b107a694ec
878c289076330178393ea49726d7cd0aba00dbac
"Support va_arg properly\n\nIt has a type argument: preserve that via explicitly supporting\n__built(...TRUNCATED)
medium
17
python
[ "pycparser/c_lexer.py", "pycparser/c_parser.py" ]
BSD-3-Clause
tasks/eliben__pycparser__878c289076/LICENSE
LICENSE
0a9e78156ae8b96caf1b4f11ac2df424b83de29736f9e757630548a6272fdf90
[]
"## Context\n\nThis repository is pycparser: a pure-Python C parser. The change makes `__builtin_va_(...TRUNCATED)
tasks/eliben__pycparser__878c289076/spec.md
false
tests/test_c_parser.py
true
[ "tests/test_c_generator.py", "tests/test_c_parser.py", "tests/test_general.py" ]
[{"path":"tasks/eliben__pycparser__878c289076/hidden/tests/test_c_generator.py","repo_path":"tests/t(...TRUNCATED)
["tests/test_c_generator.py::TestCtoC::test_builtin_va_arg","tests/test_c_parser.py::TestCParser_fun(...TRUNCATED)
["tests/test_c_generator.py::TestFunctionDeclGeneration::test_partial_funcdecl_generation","tests/te(...TRUNCATED)
true
"--- a/pycparser/c_lexer.py\n+++ b/pycparser/c_lexer.py\n@@ -517,6 +517,7 @@\n \"VOLATILE\",\n (...TRUNCATED)
tasks/eliben__pycparser__878c289076/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{"seed":1,"passed_spec_tests":true,"spec_tests":8,"passed_upstream_tests":true,"tool_calls":31,"wall(...TRUNCATED)
{"path":"tasks/eliben__pycparser__878c289076/repo.tar.gz","files":212,"bytes":1578223,"compressed_by(...TRUNCATED)
eliben__pycparser__8fd2299bef
eliben/pycparser
https://github.com/eliben/pycparser
aa32af5fb29f0a1df7266be1203afc6c283c1c3a
8fd2299bef4c23f47430e0f48773f231ccc44b69
Add ArrayDesignator AST node to distinguish array indices in designators Fixes #466
medium
29
python
[ "pycparser/c_ast.py", "pycparser/c_generator.py", "pycparser/c_parser.py" ]
BSD-3-Clause
tasks/eliben__pycparser__8fd2299bef/LICENSE
LICENSE
0a9e78156ae8b96caf1b4f11ac2df424b83de29736f9e757630548a6272fdf90
[]
"## Context\n\n`pycparser/c_ast.py` defines the AST node classes. `NamedInitializer` holds a designa(...TRUNCATED)
tasks/eliben__pycparser__8fd2299bef/spec.md
true
tests/test_array_designator.py
false
[ "tests/test_c_generator.py", "tests/test_c_parser.py" ]
[{"path":"tasks/eliben__pycparser__8fd2299bef/hidden/tests/test_c_generator.py","repo_path":"tests/t(...TRUNCATED)
["tests/test_c_generator.py::TestCtoC::test_field_and_array_designators","tests/test_c_parser.py::Te(...TRUNCATED)
["tests/test_c_generator.py::TestFunctionDeclGeneration::test_partial_funcdecl_generation","tests/te(...TRUNCATED)
false
null
null
null
null
{"path":"tasks/eliben__pycparser__8fd2299bef/repo.tar.gz","files":211,"bytes":1394409,"compressed_by(...TRUNCATED)
fsspec__filesystem_spec__4242655adc
fsspec/filesystem_spec
https://github.com/fsspec/filesystem_spec
5fc8bfb335474b6caca20beb82cd73341abb52dc
4242655adcd080e6b27f48613ec9ca1d3e22cf71
Close owned files when context entry or cleanup fails (#2392)
medium
64
python
[ "fsspec/core.py" ]
BSD-3-Clause
tasks/fsspec__filesystem_spec__4242655adc/LICENSE
LICENSE
2dc35496ce53a7307915c008a844aad53e772b49b34cdd00445067691407ec94
[]
"# Build spec: close owned files when context entry or cleanup fails\n\n## Context\n\n`fsspec/core.p(...TRUNCATED)
tasks/fsspec__filesystem_spec__4242655adc/spec.md
true
fsspec/tests/test_core.py
true
[ "fsspec/tests/test_core.py" ]
[{"path":"tasks/fsspec__filesystem_spec__4242655adc/hidden/fsspec/tests/test_core.py","repo_path":"f(...TRUNCATED)
["fsspec/tests/test_core.py::test_openfile_failed_text_wrapper_closes_file","fsspec/tests/test_core.(...TRUNCATED)
["fsspec/tests/test_core.py::test_expand_paths","fsspec/tests/test_core.py::test_expand_paths_if_nee(...TRUNCATED)
true
"--- a/fsspec/core.py\n+++ b/fsspec/core.py\n@@ -112,17 +112,21 @@\n \n self.fobjects = [f]\(...TRUNCATED)
tasks/fsspec__filesystem_spec__4242655adc/trained.patch
Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)
{"seed":2,"passed_spec_tests":true,"spec_tests":4,"passed_upstream_tests":false,"tool_calls":14,"wal(...TRUNCATED)
{"path":"tasks/fsspec__filesystem_spec__4242655adc/repo.tar.gz","files":166,"bytes":9377778,"compres(...TRUNCATED)
End of preview. Expand in Data Studio

build-spec-tasks

119 Python coding tasks from 50 open-source repositories, each a real merged change replayed from the commit before it. A task gives an agent the repository at that commit and a detailed build spec; the answer is a change to the repository, scored by tests. These are the training tasks of the Kerberos 35B-A3B 0.1 local coding model. The tasks Kerberos is evaluated on are not here.

What a task is

A merged commit that changed source code and added or edited tests, from a repository under a permissive licence. The repository is reset to the commit's parent (base_commit). The change's description is the request, a build spec describes the change in detail, and the tests the change added or edited are the hidden tests. The agent works in the whole repository: it has to find its way around, follow the project's conventions and change one or more files.

  • Sizes: 74 small, 39 medium, 6 large (by lines changed in source).
  • Licences: MIT (54), BSD-3-Clause (41), Apache-2.0 (19), BSD-2-Clause (5).
  • Up to 8 tasks from one repository.

Files

data/train.jsonl          one row per task (fields below)
tasks/<id>/spec.md        the build spec
tasks/<id>/repo.tar.gz    the repository at base_commit (git archive), unpacks to repo/
tasks/<id>/hidden/        the project's own test files for the change, as of the change's commit
tasks/<id>/LICENSE        the repository's licence file
tasks/<id>/trained.patch  the trained answer, where there is one
LICENSES/                 the licence of each part, and the MIT text for ours
scan-report.json          what the release scan checked and found

Fields

Field Meaning
id <owner>__<name>__<first 10 characters of the change's commit>
repository, repository_url the source repository
base_commit the commit the task starts from (the change's parent)
upstream_commit the real change
request the change's own commit message (co-author and sign-off lines removed)
size, changed_lines, source_files, language the real change's size and the source files it touched
licence, licence_file, licence_source_file, licence_sha256, other_licences the repository's licence (SPDX), read from its own files
spec, spec_path the build spec, inline and as a file
spec_rewritten the spec failed a check once and was rewritten by the same model
spec_tests_file where the spec's test code goes
spec_tests_pass_with_upstream_change the spec's tests also pass on the real change (69 of 119)
upstream_test_files, hidden_tests the project's test files for the change; content inline when under 64 KB, else read the file at path
changed_tests, unchanged_tests pytest node ids the change added or edited (they check the change), and the ones it left alone (regression guard)
has_trained_patch, trained_patch, trained_patch_path the answer trained on, as a unified diff applying with patch -p1 in repo/
trained_patch_model the model that wrote it
trained_result that answer's scores and cost: spec tests passed, project tests passed, tool calls, seconds, seed
snapshot repo.tar.gz: files, bytes, SHA-256, and any file left out (binary files over 5 MB: 3 in all)

How the specs were written

GLM 5.3 wrote every spec to a fixed build-spec format (context, the change file by file, the test code to add), from the request and context drawn from the repository at the base commit: the file tree, the README's start, the files defining names the request mentions, and, as hints of where and never what, the names of the functions the change edits and the new names its tests import. It was not shown the real change or the hidden tests. A spec was kept only when it passed objective checks: the structure is complete, every path and name it cites exists in the repository or is created by the spec, it repeats little of the hidden tests or the real change (measured, with a limit), and its own test code loads and fails on the base commit. A spec that failed went back to GLM 5.3 for one rewrite. Passing on the real change was not required: a spec can be right and differ from what the project did, and the spec's tests score faithfulness to the spec.

The trained answers

75 tasks carry the answer Kerberos 0.1 was trained on: the base model's own attempt that passed the spec's tests, one per task (Qwen3.6-35B-A3B, the base model before any training (UD-IQ3_XXS GGUF quant)), run by an agent harness (Warden's headless agent) with shell and file-editing tools. 49 of these 75 also pass the project's own tests. Files the test runs generated (coverage data, test caches) were removed from the diffs, and paths rewritten to a/ and b/. The other 44 tasks had no spec-passing attempt in that run.

Scoring an answer

Unpack repo.tar.gz, let the agent work in repo/ from spec.md (and the request), and take its change as a diff. Then:

  • Spec pass: add the spec's test code where the spec says (spec_tests_file), run it, and check that the tests in unchanged_tests still pass.
  • Upstream pass: copy hidden/ over the agent's tree and run the tests in changed_tests, then the regression guard.

Run tests in a sandbox with no network; the repositories' declared dependencies are needed. A scorer with a Docker image per task comes with the Kerberos public benchmark, published separately.

Licences

Each task keeps its repository's licence: the snapshot, the hidden tests and the trained patch are under the licence in tasks/<id>/LICENSE (MIT, BSD, Apache-2.0, ISC or PSF only; a task with any restrictive or unrecognised licence file in its snapshot is left out). The specs, the rows' own fields and the dataset's structure are under the MIT licence (LICENSES/MIT.txt). LICENSES/README.md lists every task's licence.

What is not included

  • The held-back evaluation tasks, any task from a repository with a held-back task, and their specs.
  • Agent traces (the conversations of the attempts), system prompts and tool definitions.
  • Task content written by a closed model: specs come only from GLM 5.3, answers only from the base model, code and tests only from the repositories.
  • Tasks left out from the training list: Legrandin__pycryptodome__9acfdf3ca0: licence is partly a public domain dedication (Unlicense wording), not on the allowed list.

Size

119 tasks; snapshots 602 MB unpacked, 213 MB compressed.

Downloads last month
3