Dataset Viewer
Auto-converted to Parquet Duplicate
messages
stringlengths
9.45k
797k
instance_id
stringlengths
45
45
rollout_patch
stringlengths
0
11.9k
target_patch
stringclasses
1 value
func_name
stringlengths
4
51
func_path
stringlengths
18
42
problem_statement
listlengths
1
1
docker_image
stringclasses
1 value
source
stringclasses
1 value
run
stringclasses
1 value
thinking_enabled
bool
1 class
exit_status
stringclasses
5 values
is_good_patch
bool
2 classes
num_agent_steps
int64
3
116
[{"role": "system", "content": "You are a helpful assistant that can interact with a computer to solve tasks.", "agent": "main", "message_type": "system_prompt"}, {"role": "user", "content": [{"type": "text", "text": "<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code repository in the directory...
sweagent-600-think__django_django_f8fab_10002
diff --git a/django/contrib/admin/actions.py b/django/contrib/admin/actions.py index 6fdf070694..4f293d185c 100644 --- a/django/contrib/admin/actions.py +++ b/django/contrib/admin/actions.py @@ -46,7 +46,7 @@ def delete_selected(modeladmin, request, queryset): for obj in queryset: obj_disp...
ModelAdmin.delete_model
django/contrib/admin/options.py
[ { "type": "text", "text": "<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code repository in the directory /testbed. Consider the following PR description:\n\n<pr_description>\nThere appears to be an issue with code abstraction and interface design related to ModelAdmin.delete_model i...
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
29
[{"role": "system", "content": "You are a helpful assistant that can interact with a computer to solve tasks.", "agent": "main", "message_type": "system_prompt"}, {"role": "user", "content": [{"type": "text", "text": "<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code repository in the directory...
sweagent-600-think__django_django_f8fab_10004
lock
django/core/files/locks.py
[ { "type": "text", "text": "<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code repository in the directory /testbed. Consider the following PR description:\n\n<pr_description>\nThere appears to be an issue with code backwards compatibility and version migration related to lock in djan...
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
null
3
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10008
"diff --git a/django/forms/boundfield.py b/django/forms/boundfield.py\nindex b6be395a67..77944fd059 (...TRUNCATED)
DayMixin.get_next_day
django/views/generic/dates.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted (exit_cost)
true
116
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10011
"diff --git a/django/test/runner.py b/django/test/runner.py\nindex 8b4c608c08..e8b38cd7e0 100644\n--(...TRUNCATED)
RemoteTestResult.addError
django/test/runner.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
89
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10012
"diff --git a/django/contrib/sites/models.py b/django/contrib/sites/models.py\nindex 744ccd358c..4b2(...TRUNCATED)
File._get_size
django/core/files/base.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
98
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10013
"diff --git a/django/utils/translation/__init__.py b/django/utils/translation/__init__.py\nindex 39d(...TRUNCATED)
_string_concat
django/utils/translation/__init__.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
39
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10014
"diff --git a/django/utils/timezone.py b/django/utils/timezone.py\nindex 29885caac8..094666c860 1006(...TRUNCATED)
override.__enter__
django/utils/timezone.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
66
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10015
"diff --git a/django/views/generic/dates.py b/django/views/generic/dates.py\nindex 938c538535..c678b(...TRUNCATED)
ImageFile.width
django/core/files/images.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted
true
81
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10017
"diff --git a/django/db/migrations/autodetector.py b/django/db/migrations/autodetector.py\nindex b2e(...TRUNCATED)
HttpResponseBase._convert_to_charset
django/http/response.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
submitted (exit_cost)
true
115
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that can interact with a compute(...TRUNCATED)
sweagent-600-think__django_django_f8fab_10018
"diff --git a/django/utils/functional.py b/django/utils/functional.py\nindex ee301879df..4cbbb6a28a (...TRUNCATED)
NestedObjects.nested
django/contrib/admin/utils.py
[{"type":"text","text":"<uploaded_files>\n/testbed\n</uploaded_files>\nI've uploaded a python code r(...TRUNCATED)
swebench/sweb.eval.x86_64.django_1776_django-7530
sweagent-600-think
django_sweagent_600_think
true
null
76
End of preview. Expand in Data Studio

SERA GLM-5.2 Django SWE-Agent - RAW T1 (first rollout, thinking enabled)

258 raw, pre-postprocess first-rollout trajectories with native GLM-5.2 reasoning traces, from the SERA SVG pipeline (paper).

Released raw so you can choose your own filtering, verification threshold and reasoning-trace handling. Companion: SERA-GLM5.2-Django-SWEAgent-Raw-T2.

Schema

Mirrors allenai/Sera-*-T1/T2:

column notes
messages JSON string - apply json.loads(). Raw SWE-agent history: role, content, thought, action, tool_calls, thinking_blocks
instance_id <source>__<django instance>
rollout_patch patch produced by this rollout
target_patch T2 only: the first-rollout patch used as the soft-verification reference
func_name, func_path the sampled starting function
problem_statement T1: the vague bug prompt. T2: the synthetic PR issue
docker_image SWE-bench container
exit_status, is_good_patch, num_agent_steps filtering aids (is_good_patch = the T1 LLM grader verdict)

Reasoning traces

Generated with GLM-5.2 extended thinking enabled (thinking: {type: enabled, budget_tokens: 8000}, temperature: 1.0). Unlike the official SERA releases - where the teacher wrote its reasoning inline and content already contains <think>...</think> - here the two channels are separate:

  • content - the short outward-facing line ("I'll start by examining...")
  • thinking_blocks - the actual deliberation (median ~1.1k chars, up to ~47k)

Wrapped into <think> tags, these traces are richer than the official release:

data median p90 max
official SERA-4.6-Lite-Best-Subset 124 1006 3999
this dataset 159 3217 39890

Caveat - empty turns. GLM only deliberates when a step is non-trivial; ~21% of assistant turns carry a thinking block, and ~25% of turns have neither text nor thinking (just a tool call). A naive add_think wrapper turns those into hollow <think></think>, which teaches the student to skip reasoning. Either drop the tags on such turns or backfill from action. (This artifact is not specific to thinking mode - it is present in non-thinking SERA-style data too.)

Turning this into training data

SERA's postprocess (sera/datagen/data/postprocess/) converts these to hermes/xml chat records, but out of the box it reads only content - to use the native traces, prefer thinking_blocks when populated. Soft verification is recomputable at any threshold from rollout_patch vs target_patch (line-level recall), so no regeneration is needed to change r.

Generation config

Teacher GLM-5.2 via an Anthropic-Messages gateway; harness SWE-agent (str_replace_editor, bash, submit); SWE-bench Django container django__django-7530 (base f8fab6f9); 115-step cap; SERA soft-verified generation (SVG). Non-thinking companion runs will be released separately.

Downloads last month
56

Paper for thientrangngv/SERA-GLM5.2-Django-SWEAgent-Raw-T1