system stringlengths 1.76k 6.02k | prompt stringlengths 249 104k | ground_truth stringlengths 21 47.2k | source stringclasses 11
values | domain stringclasses 7
values | dataset stringclasses 11
values | source_record_index int64 2 15.3k ⌀ | source_turn_index int64 0 39 ⌀ | system_tokens int64 479 853 ⌀ | prompt_tokens int64 66 11.9k ⌀ | ground_truth_tokens int64 10 11.7k ⌀ | total_tokens int64 579 12.9k ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|
Role: You are a state-transition world model for a web search and page-browsing environment. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description. Apply the ... | **Turn 1 User**
###[Goal]
According to '28 Best Academic Search Engines That make your research easier 2025', how many journal article citations does MEDLINE provide access to? You should follow the format instruction in the requestion strictly and wrap the final answer in \boxed{}.
###[Action]
Use `browser.search` to ... | ###[Observation]
L627: #26. MEDLINE®
L630: MEDLINE® is a paid subscription database for life sciences and biomedicine that includes more than 28 million citations to journal articles. For finding reliable, carefully chosen health information, Medline Plus provides a powerful search tool and even a dictionary.
L645: MED... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a web search and page-browsing environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current ... | **Turn 1 User**
###[Goal]
Tim Boyle made his debut by taking snaps for the Green Bay Packers during the 2019 NFL regular season, and completed his first career NFL pass in a subsequent game that season. State the date and opponent for each of these two milestones, in order: first snap, then first completed pass. Provid... | ###[Observation]
Search results for: pro football reference 2019 green bay packers schedule
【0†Doc 2954792】 http://www.draftdashboard.com/?hop=secnet
Instantly Optimize your Daily Fantasy Football Lineups! Our Daily Fantasy Football Tools show players with the best value and defensive matchups. Save time, get 24/7 liv... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Function as the environment model for a software-engineering shell testbed. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the suppl... | **Turn 1 User**
###[Task Description]
CloudWatch PutMetricData body compression in botocore 1.34.40+
When running the PutMetricData API call through a boto mock test when the body is gzip compressed a 404 response is returned from the moto mock which raises a botocore.errorfactory.ResourceNotFoundException exception. ... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "__init__.py\nacm\nacmpca\namp\napigateway\napigatewaymanagementapi\napigatewayv2\nappconfig\napplicationautoscaling\nappsync\nathena\nautoscaling\nawslambda\nawslambda_simple\nbackend_index.py\nbackends.py\nbackup\nbatch\nbatch_simple\nbudgets\nce\nclo... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Serve as a language world model for a web search and page-browsing environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state s... | **Turn 1 User**
###[Goal]
Find the average size of a Schapendoes.
###[Action]
Use `browser.open` to open a search result, URL, or document section for reading.
- This advances the task or changes state.
- Args: `id`="https://schapendoesclub.ca"; `num_lines`=100.
- Call:
```text
browser.open(id="https://schapendoesclub.... | ###[Observation]
Pattern not found: height | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a metadata-only MCP analytics and notebook environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do ... | **Turn 1 User**
###[Observation]
{"key": "task_00011_tren_pos_465b4962", "env_key": "carlisle"}
###[Action]
Use `use_notebook` to perform the tool action named use notebook.
- This advances the task or changes state.
- Args: `path`="extraction.ipynb"; `mode`="create".
- Call:
```text
use_notebook(path="extraction.ipynb... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"text\": \"[INFO] Connected to kernel '3994f56e-0469-4489-af3b-80e5708e144c'.\\n[INFO] Successfully activate notebook 'extraction'.\\n\\nNotebook has 1 cells.\\nShowing first 1 cells:\\n\\nIndex\\tType\\tCount\\tFirst Line\\n0\\tmarkdown\\tN/A\\tNew Note... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a web search and page-browsing environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do not add, ski... | **Turn 1 User**
###[Goal]
Who was the spouse of the first president identified as a member of the party which dominated South Carolina's state legislature in 1895? You should follow the format instruction in the requestion strictly and wrap the final answer in \boxed{}.
###[Action]
Use `browser.open` to open a search r... | ###[Observation]
Doc 2552578
URL: http://chroniclingamerica.loc.gov/lccn/sn84027008/1868-07-03/ed-1/seq-3/
L1: Newspaper Page Text
L2: . Giew? Item?.
L3: CHARLESTON, July 2.- Arrived-:
L4: Brig fjfcno, BohoOnert* Ada Richard?
L5: son, ? J. Waring, Kew Yorkj and
L6: schooner R?dichen, Boston. Sail?
L7: ed-Schbbnor Tropi... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Function as the environment model for a terminal command-execution environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the s... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a Machine Learning Engineer preparing a robust data processing and evaluation pipeline.
Your objective is to build a reproducible C++ pipeline that reads a dataset, performs data cleaning, applies bootstrap sampling, and evaluates a simple linear model.
Writ... | ###[Observation]
> > > > > Total data rows: 200
Rows with |y| > 50 (outliers): 10
Cleaned rows: 190
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a browser accessibility environment with reference IDs. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct ... | **Turn 1 User**
###[Observation]
- RootWebArea "* Job Offer Acceptance | Fastmail" [ref=e1]
- button "Mail webtasks@fastmail.com" [ref=e2]
- graphics-symbol [ref=e3]
- graphics-symbol [ref=e4]
- graphics-symbol [ref=e5]
- graphics-symbol [ref=e6]
- StaticText "Mail" [ref=e7]
- StaticText "webt... | ###[Observation]
- RootWebArea "* Job Offer Acceptance | Fastmail" [ref=e1]
- button "Mail webtasks@fastmail.com" [ref=e2]
- graphics-symbol [ref=e3]
- graphics-symbol [ref=e4]
- graphics-symbol [ref=e5]
- graphics-symbol [ref=e6]
- StaticText "Mail" [ref=e7]
- StaticText "webtasks@fastmail.co... | SFT/web/web__weblinx_browsergym | web | web__weblinx_browsergym | 14,341 | 11 | 622 | 1,555 | 1,497 | 3,674 |
Role: You are a state-transition world model for a browser accessibility environment with reference IDs. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description... | **Turn 1 User**
###[Observation]
- RootWebArea "Action Capture" [ref=e1]
- heading "Action Capture" [ref=e2]
- StaticText "Action Capture" [ref=e3]
- paragraph [ref=e4]
- StaticText "Wait for user instructions, then begin by typing a URL into a search bar." [ref=e5]
- paragraph [ref=e6]
- StaticText "... | ###[Observation]
- RootWebArea "Wattpad - Where stories live" [ref=e1]
- dialog [ref=e2]
- LayoutTable [ref=e3]
- banner [ref=e4]
- navigation [ref=e5]
- list [ref=e6]
- listitem [ref=e7]
- ListMarker "• " [ref=e8]
- link "Home" [ref=e9]
- StaticText... | SFT/web/web__weblinx_browsergym | web | web__weblinx_browsergym | 488 | 0 | 604 | 154 | 4,007 | 4,765 |
Role: Function as the environment model for a terminal command-execution environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the s... | **Turn 1 User**
###[Goal]
Please solve this task:
I am a researcher running simulations for a new spectroscopy experiment. I have a multi-service pipeline for processing the experimental data, but the core processing component is missing, and the pipeline needs to be orchestrated.
Here is the setup:
- A simulated ins... | ###[Observation]
/usr/bin/python3
Python 3.10.12
matplotlib 3.10.9
numpy 2.2.6
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Function as the environment model for a terminal command-execution environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the s... | **Turn 1 User**
###[Goal]
Please solve this task:
Wake up! It's 3 AM and you are on-call. The nightly risk calculation engine just crashed in production, halting all downstream financial reporting.
We have managed to grab a raw memory dump (`/home/user/core.dump`) and the daily input data (`/home/user/data.csv`), alo... | ###[Observation]
/home/user/calc_engine.c: In function ‘main’:
/home/user/calc_engine.c:26:74: warning: ‘%s’ directive output may be truncated writing up to 63 bytes into a region of size 47 []8;;https://gcc.gnu.org/onlinedocs/gcc/Warning-Options.html#index-Wformat-truncation=-Wformat-truncation=]8;;]
26 | ... | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a mobile accessibility-control environment. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, latest `###[Recent Observation]`, current `###[Action]`, and earlier turns as the ... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=58, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.coffeebeanventures.easyvoicerecorder', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=58
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg... | ###[Observation]
Android '', b=555,995,1070,1944, win=65, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=555,995,1070,1944, pkg='com.coffeebeanventures.easyvoicerecorder', actions='4,8,64,16908342', child_ids='1', draw=0, win=65
node=2 group '', raw_class='FrameLayout', b=555,995,1070,1944, pkg='com.cof... | SFT/android/AndroidControl | android | AndroidControl | 10,152 | 0 | 518 | 6,623 | 1,858 | 8,999 |
Role: You are a post-action environment simulator for a web search and page-browsing environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do not add, ski... | **Turn 1 User**
###[Goal]
Identify the unofficial university message board that caused a media scandal in Raleigh due to false weather closings being broadcast on a local news channel—the same news channel that once employed a UNC Chapel Hill graduate as an anchor and is headquartered near Atlantic Avenue. Crucially, t... | ###[Observation]
Doc 1518219
URL: https://www.forumotion.net/tag/relevant
L1: Free forum : A forum where resistance leaders plan their attacks and discuss relevant issues.
L2: This is the kitchen for all educational news. gists, updates, gossips, highlights, hangouts, hookups and all relevant achievements.
L3: educatio... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Operate as a language world model for a terminal command-execution environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. The ... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a data analyst tasked with processing employee interaction graphs.
Write a Python script at `/home/user/graph_analyzer.py` that processes a graph of interactions provided in a CSV file and outputs specific node metrics in JSON format.
Your script must accept ... | ###[Observation]
source,target,relation,weight
A,B,manager,10
A,C,peer,5
B,C,manager,8
C,A,report,3
D,B,manager,10
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Serve as a language world model for a mobile accessibility-control environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state s... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=2773, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.flyersoft.moonreader', actions='4,8,64,16908342', child_ids='1', draw=0, win=2773
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='com.flyersoft.... | ###[Observation]
Android '', b=0,0,1080,2400, win=2773, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.flyersoft.moonreader', actions='4,8,64,16908342', child_ids='1', draw=0, win=2773
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='com.flyersoft.moonreader', act... | SFT/android/AndroidControl | android | AndroidControl | 967 | 0 | 506 | 1,163 | 1,112 | 2,781 |
Role: Function as the environment model for a software-engineering shell testbed. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the suppl... | **Turn 1 User**
###[Task Description]
iot: actions where ids are sent in the body won't work when they have URL reserved characters
The root cause is when we use characters that need to be url-encoded on ids (e.g. `thingGroupName`). When those are created the id is part of the URL so moto creates them url-encoded in it... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "/testbed/moto/iot/utils.py\n/testbed/moto/iot/responses.py\n/testbed/moto/iot/models.py\n/testbed/moto/iot/__init__.py\n/testbed/moto/iot/exceptions.py\n/testbed/moto/iot/urls.py\n", "stderr": "", "duration_ms": 31} | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Serve as a language world model for a web search and page-browsing environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state s... | **Turn 1 User**
###[Goal]
Find the percentage of executives in the Accenture Technology Trends 2025 | Technology Vision report who stated that AI agents would improve the flexibility of their organization’s digital architecture. Present the exact percentage only.
###[Action]
Use `browser.search` to search the web or re... | ###[Observation]
Search results for: "Accenture" "Technology Vision 2025" AI agents flexibility digital architecture survey executives
【0†Doc 2501999】 http://aiplanner.launchconsulting.com/
This tool will help you assess an opportunity to apply AI in your company, and provide a high level business case and plan for do... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Act as a language world model for a software-engineering shell testbed. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the provided `###[Task Description]`, the current `###[Action]`, and prior observat... | **Turn 1 User**
###[Task Description]
QST: pd.series indexing True auto 0 change
### Research
- [X] I have searched the [[pandas] tag](https://stackoverflow.com/questions/tagged/pandas) on StackOverflow for similar questions.
- [X] I have asked my usage related question on [StackOverflow](https://stackoverflow.com).
... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "commit 2a6c6a2d131e49aa1b0272d5d5bef94c9810f67f\nAuthor: Patrick Hoefler <61934744+phofl@users.noreply.github.com>\nDate: Tue Sep 27 00:58:09 2022 +0200\n\n BUG: Series.getitem not falling back to positional for bool index (#48662)\n \n * BU... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Act as a language world model for a metadata-only MCP analytics and notebook environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, an... | **Turn 1 User**
###[Observation]
{"key": "task_00036_comp_dw", "env_key": "carlisle"}
###[Action]
Use `list_warehouses` to retrieve or inspect list warehouses information.
- This advances the task or changes state.
- Call:
```text
list_warehouses()
```
**Turn 1 Assistant**
###[Observation]
{"type": "text", "text": "{\n... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": false,\n \"warehouse\": \"dw\",\n \"columns\": [],\n \"rows\": [],\n \"row_count\": 0,\n \"execution_time_ms\": 1.2,\n \"error\": \"Table or column not found: Catalog Error: Scalar Function with name t_cdf does not exist!\\nDid you mean \\\"gcd\\\"?\\n\\nLIN... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Operate as a language world model for a mobile accessibility-control environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1440,3120, win=74, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1440,3120, pkg='com.tul.tatacliq', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=74
node=2 group '', raw_class='LinearLayout', b=0,0,1440,3036, pkg='com.tul.tatacliq', act... | ###[Observation]
Android '', b=0,0,1440,3120, win=57, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1440,3120, pkg='com.tul.tatacliq', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=57
node=2 group '', raw_class='LinearLayout', b=0,0,1440,3036, pkg='com.tul.tatacliq', actions='4,8,64,169... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a terminal command-execution environment. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, latest `###[Observation]`, current `###[Action]`, and earlier turns as the evidence ... | **Turn 1 User**
###[Goal]
Please solve this task:
You are helping a technical writer automate their documentation build pipeline. They have a massive monolithic text file that needs to be split, converted to basic HTML, and safely merged into a single output file using a concurrent background worker.
Your objective i... | ###[Observation]
Watcher started with PID: 482
Expected 3 files to be processed
All files processed!
Stopping watcher (PID: 482)...
Watcher stopped.
Pipeline complete!
Output file: /home/user/build//output.html
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a browser accessibility-control environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Recent Observation]`, action text, and history as state evidence. Do not ... | **Turn 1 User**
###[Observation]
Web 'One Stop Market', focused
[213] generic, busy=1
[222] banner '', visible
[225] list '', visible
[226] listitem '', visible
[227] link 'My Account', clickable, visible
[228] listitem '', visible
[229] link 'My Wish List', clickable, visible
[230] listitem... | ###[Observation]
Web 'Patio, Lawn & Garden', focused
[224] banner '', visible
[227] list '', visible
[228] listitem '', visible
[229] link 'My Account', clickable, visible
[230] listitem '', visible
[231] link 'My Wish List 114 items', clickable, visible
StaticText '114 items'
[233] listitem ''... | SFT/web/A3-Synth | web | A3-Synth | 1,627 | 0 | 500 | 3,582 | 4,778 | 8,860 |
Role: Function as the environment model for a terminal command-execution environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the s... | **Turn 1 User**
###[Goal]
Please solve this task:
You are an AI assistant helping a researcher organize and analyze a massive dataset of academic citations.
The researcher has an SQLite database located at `/home/user/research.db`. It contains a hierarchical knowledge graph of papers and citations with the following... | ###[Observation]
100000
296794
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Act as a language world model for a web search and page-browsing environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, and prior obse... | **Turn 1 User**
###[Goal]
在一款2019年上线、以‘万里独行’为外号的高手作为最终BOSS的日常组队副本中，玩家需先清理包含两波不同小怪和两批其他怪物，最后挑战BOSS。请计算该副本挑战中，在见到最终BOSS之前需要击败的所有怪物总数（包括杂兵与分批怪物），不包括BOSS本身，仅输出总数答案。 You should follow the format instruction in the requestion strictly and wrap the final answer in \boxed{}.
###[Action]
Use `browser.search` to search the web o... | ###[Observation]
Search results for: "万里独行" 角色 游戏 外号
【0†Doc 12650332】 https://www.webnovel.com/book/8335460805000105/22427064367890940
Translator: EndlessFantasy Translation Editor: EndlessFantasy Translation
Prologue: Would You Like To Be A Dimensional-Traveler?
"I have finally gotten the nickname of the Ultimate Afr... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a state-transition world model for a web search and page-browsing environment. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description. Apply the ... | **Turn 1 User**
###[Goal]
Identify the Himalayan Parmeliaceae genus with olive brown coloration, effigurate pseudocyphellae, and bifusiform conidia that uses Anaptychia ciliaris as a substrate.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant documents.
- This advances the task or chan... | ###[Observation]
Search results for: Phacopsis Anaptychia ciliaris lichenicolous Parmeliaceae
【0†Doc 10900628】 https://dryades.units.it/home/index.php?procedure=ext_key_home&key_type=lic&key_id=1882
KEYS TO THE LICHENS OF ITALY - 22) GYALECTOID LICHENS (Coenogonium, Cryptodiscus, Diploschistella, Geisleria, Gyalecta, ... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Function as the environment model for a browser accessibility environment with reference IDs. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction sta... | **Turn 1 User**
###[Observation]
- RootWebArea "Google Calendar - Friday, June 9, 2023, today" [ref=e1]
- button "Skip to main content" [ref=e2]
- button "Keyboard shortcuts" [ref=e3]
- button "Accessibility Feedback" [ref=e4]
- banner [ref=e5]
- button "Main drawer" [ref=e6]
- graphics-symbol [ref=e7... | ###[Observation]
- RootWebArea "Google Calendar - Friday, June 9, 2023, today" [ref=e1]
- button "Skip to main content" [ref=e2]
- button "Keyboard shortcuts" [ref=e3]
- button "Accessibility Feedback" [ref=e4]
- banner [ref=e5]
- button "Main drawer" [ref=e6]
- graphics-symbol [ref=e7]
- generic ... | SFT/web/web__weblinx_browsergym | web | web__weblinx_browsergym | null | null | null | null | null | null |
Role: You are a language world model for a metadata-only MCP analytics and notebook environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Acti... | **Turn 1 User**
###[Observation]
{"key": "task_00007_tren_pos_v4", "env_key": "carlisle"}
###[Action]
Use `list_warehouses` to retrieve or inspect list warehouses information.
- This advances the task or changes state.
- Call:
```text
list_warehouses()
```
**Turn 1 Assistant**
###[Observation]
{"type": "text", "text": ... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"warehouse\": \"pos\",\n \"table\": {\n \"name\": \"fct_transactions\",\n \"schema\": \"default\",\n \"columns\": [\n {\n \"name\": \"transaction_id\",\n \"type\": \"TEXT\",\n \"nullable\": true\n },\n {\n ... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of an MCP business-tool environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current state. App... | **Turn 1 User**
###[Goal]
For the existing app with bundle ID 'com.acme.habitforge', create a new iOS app version '1.1' in status 'prepare for submission' and set its 'What’s New' text to 'Added dark mode, improved reminders, and fixed notification bugs.'
###[Action]
Use `list_tools` to list available MCP tools before ... | ###[Observation]
Available MCP Tools (29 tools):
================================================================================
1. mcp_tool_mcp_server_create_app
Description: Create a new app
Create a new app owned by the fixed current user (user_id=1).
### Responses:
**200**: Successful Response (Succes... | SFT/mcp/AWM | mcp | AWM | null | null | null | null | null | null |
Role: You are a state-transition world model for a terminal command-execution environment. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description. Apply the re... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a systems programmer debugging a critical application. We have an executable `/home/user/app` that depends on a shared library `/home/user/libcore.so`.
Currently, running `/home/user/app` fails at runtime because `libcore.so` relies on an undefined external sy... | ###[Observation]
(no output)
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are a deterministic language world model for a metadata-only MCP analytics and notebook environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state rec... | **Turn 1 User**
###[Observation]
{"key": "task_00424_data_dw_a1701a1b", "env_key": "carlisle"}
###[Action]
Use `list_warehouses` to retrieve or inspect list warehouses information.
- This advances the task or changes state.
- Call:
```text
list_warehouses()
```
**Turn 1 Assistant**
###[Observation]
{"type": "text", "te... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"warehouse\": \"dw\",\n \"table\": \"dim_users\",\n \"s3_path\": \"s3://workspace/imputation/dim_users_imputed.parquet\",\n \"rows_written\": 5495,\n \"columns\": [\n \"user_id\",\n \"email\",\n \"display_name\",\n \"user_type\",\n \"a... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Act as a language world model for a software-engineering shell testbed. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the provided `###[Task Description]`, the current `###[Action]`, and prior observat... | **Turn 1 User**
###[Task Description]
RandAffined fails with KeyError when allow_missing_keys and key at index 0 is missing
**Describe the bug**
We have a pipeline in which not all samples have the same keys.
The `RandAffined` transform accesses fallback data via index 0.
If index 0 is a missing key, then this trans... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "tests/test_save_imaged.py\ntests/test_check_missing_files.py\n", "stderr": "", "duration_ms": 115} | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a language world model for a terminal command-execution environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and any ... | **Turn 1 User**
###[Goal]
Please solve this task:
You are an incident responder investigating a suspected breach on a Linux web server. The server runs a custom authentication service written in C that utilizes JSON Web Tokens (JWT).
We suspect the attacker exploited a vulnerability in the JWT implementation and subs... | ###[Observation]
total 1364
drwxrwxrwx 2 root root 35 Apr 30 03:16 .
drwxr-xr-x 5 root root 124 Apr 30 03:16 ..
-rwsr-xr-x 1 root root 1396520 Apr 30 03:16 system_helper
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are a language world model for a mobile accessibility-control environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=2220, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.conceptivapps.blossom', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=2220
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='com.conce... | ###[Observation]
Android '', b=0,0,1080,2400, win=2220, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.conceptivapps.blossom', actions='4,8,64,16908342', child_ids='1,2,4', draw=0, win=2220
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='com.conceptivapps.blossom... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: Operate as a language world model for a software-engineering shell testbed. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. The curr... | **Turn 1 User**
###[Task Description]
CSVIterableDataset doesn't work in the same way as CSVDataset
**Is your feature request related to a problem? Please describe.**
Here is the feedback from @holgerroth :
When I am using `CSVIterableDataset`, I get this error
```
Traceback (most recent call last):
File "run_... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "78: \"convert_tables_to_dicts\",\n1082:def convert_tables_to_dicts(\n", "stderr": "", "duration_ms": 69} | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a metadata-only MCP analytics and notebook environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do ... | **Turn 1 User**
###[Observation]
{"key": "task_00130_data_dw", "env_key": "carlisle"}
###[Action]
Use `request_role` to perform the tool action named request role.
- This advances the task or changes state.
- Args: `role_name`="data_engineer"; `reason`="Need access to dw warehouse to analyze experiments table for dupli... | ###[Observation]
{"type": "text", "text": "{\n \"approved\": true,\n \"assignment_id\": \"7d18b935\",\n \"role_name\": \"data_engineer\",\n \"description\": \"Full access to all warehouses including schema modifications\",\n \"warehouse_access\": [\n \"*\"\n ],\n \"permissions\": [\n \"read\",\n \"write... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Function as the environment model for a terminal command-execution environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the s... | **Turn 1 User**
###[Goal]
Please solve this task:
You are tasked with building a lightweight, Bash-based artifact manager API that curates a local binary repository. The repository is located at `/home/user/artifacts`.
Due to some bad curation scripts in the past, the artifact directory contains complex symlink stru... | ###[Observation]
[/home/user]$ [/home/user]$ bash: iptables: command not found
[/home/user]$ ---
[/home/user]$ [/home/user]$ [/home/user]$ > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > [/home/... | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Act as a language world model for a browser accessibility-control environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, and prior obs... | **Turn 1 User**
###[Observation]
Web 'Orders / Operations / Sales / Magento Admin', focused
[171] link 'Magento Admin Panel', clickable, visible
[172] image 'Magento Admin Panel', visible
[173] navigation '', visible
[174] menubar '', visible, orientation='horizontal'
[176] link '\ue604 DASHBOARD', clickable, ... | ###[Observation]
Web 'Orders / Operations / Sales / Magento Admin', focused
[171] link 'Magento Admin Panel', clickable, visible
[172] image 'Magento Admin Panel', visible
[173] navigation '', visible
[174] menubar '', visible, orientation='horizontal'
[176] link '\ue604 DASHBOARD', clickable, visible
Stat... | SFT/web/A3-Synth | web | A3-Synth | null | null | null | null | null | null |
Role: You are a language world model for a mobile accessibility-control environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=344, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.yummly.android', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=344
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.yummly.android... | ###[Observation]
Android '', b=0,0,1080,2400, win=345, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.yummly.android', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=345
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.yummly.android', actions='4,8,... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are a language world model for a browser accessibility-control environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Recent Observation]`, the current `###[Action]`... | **Turn 1 User**
###[Observation]
Web 'New Issue · Primer / design · GitLab', focused
[71] banner '', visible
[72] link 'Skip to content', clickable
StaticText 'GitLab'
[78] link 'Dashboard', clickable, visible
[79] image '', visible
[82] list '', visible
[83] listitem '', visible
[84] button '', clic... | ###[Observation]
Web 'New Issue · Primer / design · GitLab', focused
[71] banner '', visible
[72] link 'Skip to content', clickable
StaticText 'GitLab'
[78] link 'Dashboard', clickable, visible
[79] image '', visible
[82] list '', visible
[83] listitem '', visible
[84] button '', clickable, visible, ... | SFT/web/A3-Synth | web | A3-Synth | 6,905 | 4 | 612 | 2,213 | 2,159 | 4,984 |
Role: You are a language world model for a metadata-only MCP analytics and notebook environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Acti... | **Turn 1 User**
###[Observation]
{"key": "task_00009_cros_dw_7ac892c9", "env_key": "carlisle"}
###[Action]
Use `list_warehouses` to retrieve or inspect list warehouses information.
- This advances the task or changes state.
- Call:
```text
list_warehouses()
```
**Turn 1 Assistant**
###[Observation]
{"type": "text", "te... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"warehouse\": \"dw\",\n \"columns\": [\n \"null_users\",\n \"null_customers\"\n ],\n \"rows\": [\n [\n 0,\n 0\n ]\n ],\n \"row_count\": 1,\n \"execution_time_ms\": 7.31,\n \"error\": null\n}", "annotations": null, "meta": nul... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Operate as a language world model for a web search and page-browsing environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Goal]
A Norman-style chapel built during a period of global financial crisis in a western American city includes a bell inscribed with a famous quote from a Lebanese-American author whose major poetic work was published in the 1920s and who bequeathed his royalties to his Levantine birthplace. The b... | ###[Observation]
Search results for: Norman chapel built 1929 1930s Great Depression western US city bell
【0†Doc 6748717】 http://nhunited.org/history/
One of the moving spirits in the organizing of the Normal Heights Methodist Episcopal Church in 1913 was Mrs. A. J. Norton. Living on what is now Felton Street, she fel... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Serve as a language world model for a mobile accessibility-control environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state s... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=1163, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.coffeebeanventures.easyvoicerecorder', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1163
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337,... | ###[Observation]
Android '', b=0,0,1080,2400, win=1163, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.coffeebeanventures.easyvoicerecorder', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1163
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.coffee... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are a deterministic language world model for a browser accessibility-control environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply ... | **Turn 1 User**
###[Observation]
Web 'Camping | OpenStreetMap', focused
[44] banner '', visible
[45] heading 'OpenStreetMap logo OpenStreetMap', visible
[46] link 'OpenStreetMap logo OpenStreetMap', clickable, visible
[49] image 'OpenStreetMap logo', visible
StaticText ''
[51] navigation '', visible
... | ###[Observation]
Web 'Camping | OpenStreetMap', focused
[44] banner '', visible
[45] heading 'OpenStreetMap logo OpenStreetMap', visible
[46] link 'OpenStreetMap logo OpenStreetMap', clickable, visible
[49] image 'OpenStreetMap logo', visible
StaticText ''
[51] navigation '', visible
[95] link 'Edit'... | SFT/web/A3-Synth | web | A3-Synth | null | null | null | null | null | null |
Role: You are a language world model for a web search and page-browsing environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Goal]
Find a state school for the blind in the American South where its former facility was later repurposed as the state's gubernatorial residence. Focus on schools founded before 1870 that moved after the Civil War.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant... | ###[Observation]
Doc 10945140
URL: http://www.deuelarea.com/business-resources/resource-library/p/item/2827/buffalo-ridge-resort-sd-historical-societys-case-study
L1: Resource Library« Back to Search Results
L2: Featured here in the Resource Library is a project summary from the South Dakota State Historical Society. I... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a software-engineering shell testbed. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, the provided `###[Task Description]`, current `###[Action]`, and earlier turns as the ev... | **Turn 1 User**
###[Task Description]
ENH: Add SAS7BDATReader and XportReader to `pandas/api/typing/__init__.py`
### Feature Type
- [X] Adding new functionality to pandas
- [ ] Changing existing functionality in pandas
- [ ] Removing existing functionality in pandas
### Problem Description
These classes can be re... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "\"\"\"\nPublic API classes that store intermediate results useful for type-hinting.\n\"\"\"\n\nfrom pandas._libs import NaTType\nfrom pandas._libs.missing import NAType\n\nfrom pandas.core.groupby import (\n DataFrameGroupBy,\n SeriesGroupBy,\n)\... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Operate as a language world model for a web search and page-browsing environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Goal]
Find the city where Olivier-Napoléon Drouin died.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant documents.
- This advances the task or changes state.
- Args: `query`="\"Olivier-Napoléon Drouin\" \"maire de\""; `topn`=10.
- Call:
```text
browser.search(query... | ###[Observation]
Search results for: "Olivier-Napoléon Drouin" "Charlesbourg" OR "Beauport" OR "Sillery" OR "Sainte-Foy"
【0†Doc 6463538】 http://acadianroots.blogspot.com/2009/04/name-places-in-acadie.html
Trying different Fonts on my blog. This one doesn't seem much different that the one I usually use. Anyway getting... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a deterministic language world model for a mobile accessibility-control environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply t... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=2919, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.surveyheart', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=2919
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.surveyheart', a... | ###[Observation]
Android '', b=0,0,1080,2400, win=2919, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.surveyheart', actions='4,8,64,16908342', child_ids='1,2,4', draw=0, win=2919
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.surveyheart', actions='4,8,64,1... | SFT/android/AndroidControl | android | AndroidControl | 10,422 | 1 | 549 | 4,720 | 3,771 | 9,040 |
Role: You are the simulator side of a language world model for a metadata-only MCP analytics and notebook environment. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, latest `###[Observation]`, current `###[Action]`, and earlier turns as... | **Turn 1 User**
###[Observation]
{"key": "task_00329_data_dw_9fde1145", "env_key": "carlisle"}
###[Action]
Use `list_warehouses` to retrieve or inspect list warehouses information.
- This advances the task or changes state.
- Call:
```text
list_warehouses()
```
**Turn 1 Assistant**
###[Observation]
{"type": "text", "te... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"warehouse\": \"dw\",\n \"columns\": [\n \"min(created_at)\",\n \"max(created_at)\",\n \"count(DISTINCT created_at)\"\n ],\n \"rows\": [\n [\n \"2026-01-02T00:00:00+00:00\",\n \"2026-04-02T00:00:00+00:00\",\n 2\n ]\n ],\... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Act as a language world model for a metadata-only MCP analytics and notebook environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, an... | **Turn 1 User**
###[Observation]
{"key": "wallst_tmt_task_71_development_shipment_3_feb_7_v1", "env_key": "wallst"}
###[Action]
Use `bash` to execute a shell command in the terminal or task workspace.
- This advances the task or changes state.
- Args: `command`="ls Options_task_71.xlsx".
- Call:
```text
bash(command="l... | ###[Observation]
{"type": "text", "text": "{\n \"filename\": \"Options_task_71.xlsx\",\n \"sheets\": [\n {\n \"name\": \"Real option\",\n \"index\": 0,\n \"type\": \"worksheet\",\n \"dimensions\": \"A1:Z24\",\n \"rowCount\": 24,\n \"columnCount\": 26,\n \"hidden\": false,\n ... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a metadata-only MCP analytics and notebook environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct ... | **Turn 1 User**
###[Observation]
{"key": "task_00003_pand_dw_30bd06", "env_key": "carlisle"}
###[Action]
Use `use_notebook` to perform the tool action named use notebook.
- This advances the task or changes state.
- Args: `path`="customer_analytics_dashboard.ipynb"; `mode`="create".
- Call:
```text
use_notebook(path="c... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"text\": \"Extracted: customers.csv\\nExtracted: orders.csv\\n\"\n}", "annotations": null, "meta": null} | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: Operate as a language world model for a web search and page-browsing environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Goal]
Find the city where the Bank of the Orient is headquartered.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant documents.
- This advances the task or changes state.
- Args: `query`="\"Bank of the Orient\" California"; `topn`=10.
- Call:
```text
browser.search(q... | ###[Observation]
Search results for: "Bank of the Orient" US bank
【0†Doc 7215769】 http://www.chinagoabroad.com/en/contributor/banco-cathay
Banco Cathay was founded 19 years ago to break paradigms and demonstrate that business and trade development between Costa Rica and Asian countries is more than just a dream. We we... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a deterministic language world model for a web search and page-browsing environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply t... | **Turn 1 User**
###[Goal]
Identify the director of the music video for "Such Great Heights" by The Postal Service. Also verify if this director's earlier work is known for its offbeat comedic tone.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant documents.
- This advances the task or ... | ###[Observation]
https://imvdb.com/n/josh-melnick
URL: https://imvdb.com/n/josh-melnick
L1: Document not found | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a terminal command-execution environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current st... | **Turn 1 User**
###[Goal]
Please solve this task:
I need you to help me prepare some training data by reconstructing a custom feature extraction model and debugging a plotting script.
I have a dataset of raw samples in `/home/user/data/raw_samples.npy` (shape: `1000, 128`).
The custom model weights are stored in `/... | ###[Observation]
scipy version: 1.15.3
scipy.stats available
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Operate as a language world model for an MCP business-tool environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. The current ... | **Turn 1 User**
###[Goal]
In my 'Headquarters' site, create a preset named 'Cleaning Mode' that sets all zones' fixtures to 100% brightness, neutral white (4000K) or equivalent RGB color, a transition_time of 2 seconds, and save it for reuse across all buildings in this site.
###[Action]
Use `list_tools` to list availa... | ###[Observation]
{
"preset_zone_setting": {
"id": 2,
"preset_id": 1,
"zone_id": null,
"applies_to_all_zones": true,
"brightness": 100,
"color_temp": 4000,
"color_mode": "white",
"color_r": null,
"color_g": null,
"color_b": null,
"transition_time_ms": 2000,
"created_at":... | SFT/mcp/AWM | mcp | AWM | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a mobile accessibility-control environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current ... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=982, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.tul.tatacliq', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=982
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.tul.tatacliq', a... | ###[Observation]
Android '', b=0,0,1080,2400, win=1004, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.tul.tatacliq', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1004
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.tul.tatacliq', actions='4,8,64... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are a language world model for a web search and page-browsing environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Goal]
Find universities in the Mid-Eastern Athletic Conference (MEAC) that are located in cities with the same name as their state, and identify which one upset a #2 seed in its first NCAA tournament appearance in the 2010s.
###[Action]
Use `browser.search` to search the web or retrieval index for r... | ###[Observation]
Search results for: site:collegefactual.com MEAC conference
【0†Doc 5662402】 http://rushthecourt.net/2008/12/13/checking-in-on-the-meac-2/
Checking in on the… MEACPosted by rtmsf on December 13th, 2008
JC at HBCUSportsBlog is the RTC correspondent for the SWAC and MEAC Conferences.
And now its time for... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a language world model for a web search and page-browsing environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Goal]
Find information about an industrial city northeast of Damascus, Syria that had a brewery established in 2010. I need to identify which industrial city this is.
###[Action]
First, use `browser.open` to open a search result, URL, or document section for reading.
- This advances the task or chan... | ###[Observation]
Doc 7836
URL: https://en.wikipedia.org/wiki/Adra,_Syria
L1: Adra, Syria
L2:
L3: From Wikipedia, the free encyclopedia
L4:
L5: 'Adra
L6:
L7: عَدْرَا
L8: Village
L9: 'Adra
L10: Coordinates: 33°36′00″N 36°30′54″E / 33.6°N 36.515°E
L11: Country Syria
L12: Governorate Rif Dimashq
L13: D... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Act as a language world model for an MCP business-tool environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, and prior observations a... | **Turn 1 User**
###[Goal]
Create a subscriber segment named 'Highly Engaged Paid Subscribers' that includes only subscribers on a paid plan who have opened at least 60% of emails and clicked at least 20% of links in the past 90 days, and send them a targeted email campaign with the subject line 'Founders-Only Q1 AMA In... | ###[Observation]
{
"subscribers": [
{
"subscriber_id": 1,
"email": "jane.reader@example.com",
"name": "Jane Reader",
"plan_type": "free",
"billing_period": "none",
"signup_timestamp": "2026-05-19T12:15:58",
"lifetime_emails_opened": 40,
"lifetime_links_clicked": 15
... | SFT/mcp/AWM | mcp | AWM | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a browser accessibility-control environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Recent Observation]`, action text, and history as state evidence. Do not ... | **Turn 1 User**
###[Observation]
Web 'Multilingual support for prompts (#21) · Issues · Byte Blaze / empathy-prompts · GitLab', focused
[87] banner '', visible
[88] link 'Skip to content', clickable
StaticText 'GitLab'
[94] link 'Dashboard', clickable, visible
[95] image '', visible
[98] list '', visible
... | ###[Observation]
Web 'Multilingual support for prompts (#21) · Issues · Byte Blaze / empathy-prompts · GitLab', focused
[87] banner '', visible
[88] link 'Skip to content', clickable
StaticText 'GitLab'
[94] link 'Dashboard', clickable, visible
[95] image '', visible
[98] list '', visible
[99] listitem '... | SFT/web/A3-Synth | web | A3-Synth | 12,324 | 14 | 626 | 3,587 | 3,538 | 7,751 |
Role: You are a post-action environment simulator for a software-engineering shell testbed. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, the provided `###[Task Description]`, action text, and history as state evidence. Do not add... | **Turn 1 User**
###[Task Description]
plots --open: not working on macOS
We use relpath to open files in the browser.
https://github.com/iterative/dvc/blob/59cfc2c4a43e9b9d2fdb583da80d24e82ae39c26/dvc/command/plots.py#L60
This is mostly done to support `wsl`, and the fact that all other Platform seem to work, if ... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "E Args:\nE assert ('file:///tmp.../index.html',) == (PosixTmpDir(...index.html'),)\nE At index 0 diff: 'file:///tmp/pytest-of-root/pytest-1/test_plots_diff_open_failed0/dvc_plots/index.html' != PosixTmpDir('/tmp/pytest-of-root/pytes... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a state-transition world model for a terminal command-execution environment. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description. Apply the re... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a web developer working on a request validation and rate-limiting feature for a new API endpoint. Part of the validation is handled by a high-performance C binary, but the wrapper logic needs to be implemented in Bash.
First, you need to fix the build system f... | ###[Observation]
evil1 exit: 1
evil2 exit: 0
evil3 exit: 0
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Operate as a language world model for a mobile accessibility-control environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Observation]
Android '', b=28,1095,1052,1378, win=242, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=28,1095,1052,1378, pkg='com.ToDoReminder.gen', actions='4,8,64,16908342', child_ids='1', draw=0, win=242
node=2 group '', raw_class='LinearLayout', b=70,1137,1010,1336, pkg='com.ToDo... | ###[Observation]
Android '', b=28,1011,1052,1461, win=240, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=28,1011,1052,1461, pkg='com.ToDoReminder.gen', actions='4,8,64,16908342', child_ids='1', draw=0, win=240
node=2 group '', raw_class='LinearLayout', b=70,1053,1010,1419, pkg='com.ToDoReminder.gen', a... | SFT/android/AndroidControl | android | AndroidControl | 3,194 | 7 | 536 | 595 | 726 | 1,857 |
Role: You are an observation generator for a language world model of a web search and page-browsing environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current ... | **Turn 1 User**
###[Goal]
A former educational institution in Fergus Falls, Minnesota, occupies a Romanesque Revival building constructed in 1901, which later became home to a private academy owned by a church strongly emphasizing missions and a non-liturgical, evangelical tradition. The original architects of this bui... | ###[Observation]
Search results for: Fergus Falls Minnesota former school building Romanesque Revival NRHP 1901
【0†Doc 4093】 https://en.wikipedia.org/wiki/Fergus_Falls_Regional_Treatment_Center
Fergus Falls Regional Treatment Center
Fergus Falls State Hospital Complex
U.S. National Register of Historic Places
Fergus ... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a software-engineering shell testbed. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, the provided `###[Task Description]`, current `###[Action]`, and earlier turns as the ev... | **Turn 1 User**
###[Task Description]
BUG: Pandas 1.4.x - calling groupBy(...).apply(func) on an empty dataframe invokes func
As of `pandas 1.4.0`, if you call `df.groupby(['some_column']).apply(_func)` on an empty dataframe, the `func` will be called with an empty group dataframe, causing the function to potentially e... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": " def apply(self) -> DataFrame | Series:\n \"\"\"compute the results\"\"\"\n # dispatch to agg\n if is_list_like(self.f):\n return self.apply_multiple()\n\n # all empty\n if len(self.columns) == 0 and len... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a mobile accessibility-control environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Recent Observation]`, action text, and history as state evidence. Do not a... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=5651, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.microsoft.whiteboard.publicpreview', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=5651
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, p... | ###[Observation]
Android '', b=0,0,1080,2400, win=5656, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='android', actions='4,8,64,16908342', child_ids='1', draw=0, win=5656
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='android', actions='4,8,64,16908342', child_ids=... | SFT/android/AndroidControl | android | AndroidControl | 13,610 | 3 | 519 | 2,838 | 1,923 | 5,280 |
Role: You are a state-transition world model for a mobile accessibility-control environment. Your prediction is the next observation after the given action, not a recommendation or solution.
Input: Treat the goal, metadata, latest observation, action text, and earlier turns as the current state description. Apply the ... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=730, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.android.settings', actions='4,8,64,16908342', child_ids='1', draw=0, win=730
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.android.settings... | ###[Observation]
Android '', b=0,0,1080,2400, win=730, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.android.settings', actions='4,8,64,16908342', child_ids='1', draw=0, win=730
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.android.settings', actions='4,8,... | SFT/android/AndroidControl | android | AndroidControl | 7,862 | 6 | 602 | 2,890 | 2,815 | 6,307 |
Role: You are the simulator side of a language world model for a software-engineering shell testbed. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, the provided `###[Task Description]`, current `###[Action]`, and earlier turns as the ev... | **Turn 1 User**
###[Task Description]
error: Expected type in class pattern; found "Any" (cannot be ignored, error without line number)
Found a strange bug today: https://mypy-play.net/?mypy=master&python=3.11&gist=5069b99b248041428036a717e52a3476
```python
from pandas import DataFrame, Series
from typing import T... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "2d8ad8ef2 Fix async iterator body stripping (#15491)\n2f56b3f8c Unbreak CI (#15505)\na47279b86 Remove `py` and `pytest-forked` test dependencies (#15501)\nc918d40b1 Use `()` for `get_function_sig_hook` docs (#15499)\n2e304ca36 Fix PEP 561 editable inst... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a software-engineering shell testbed. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, the provided `###[Task Description]`, current `###[Action]`, and earlier turns as the ev... | **Turn 1 User**
###[Task Description]
BUG: groupby with categorical result order differs for index type
When a categorical grouper is used on a DataFrame with a range index, the sort order of the values depends only on whether the categorical is ordered. When the grouper is the index, the sort order of the values depen... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": " \"a\": Categorical([2, 1, 2, 3], categories=[1, 4, 3, 2], ordered=ordered),\n \"b\": range(4),\n }\n )\n if index_kind == \"range\":\n keys = [\"a\"]\n elif index_kind == \"single\":\n keys = [\"a\"]... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Operate as a language world model for a web search and page-browsing environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. Th... | **Turn 1 User**
###[Goal]
Find information about newly established public universities in Bihar that began operating academically in the late 2010s, particularly any that received colleges previously affiliated with Magadh University.
###[Action]
Use `browser.search` to search the web or retrieval index for relevant do... | ###[Observation]
Doc 8441
URL: https://ppup.ac.in/
L1: ::: PATLIPUTRA UNIVERSITY, PATNA :::
L2:
L3: About Us
L4:
L5: Patliputra University, Patna was established on 18th March, 2018, by the order of the Government of Bihar, vide Letter/memo no. 15/M1-71/2016-700, dated 09.04.2018. All the colleges of Patna and Nala... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a browser accessibility environment with reference IDs. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct ... | **Turn 1 User**
###[Observation]
- RootWebArea "* Compose message | Fastmail" [ref=e1]
- button "Mail webtasks@fastmail.com" [ref=e2]
- graphics-symbol [ref=e3]
- graphics-symbol [ref=e4]
- graphics-symbol [ref=e5]
- graphics-symbol [ref=e6]
- StaticText "Mail" [ref=e7]
- StaticText "webtasks@... | ###[Observation]
- RootWebArea "* Compose message | Fastmail" [ref=e1]
- button "Mail webtasks@fastmail.com" [ref=e2]
- graphics-symbol [ref=e3]
- graphics-symbol [ref=e4]
- graphics-symbol [ref=e5]
- graphics-symbol [ref=e6]
- StaticText "Mail" [ref=e7]
- StaticText "webtasks@fastmail.com" [r... | SFT/web/web__weblinx_browsergym | web | web__weblinx_browsergym | 11,656 | 12 | 679 | 3,522 | 1,587 | 5,788 |
Role: Act as a language world model for a software-engineering shell testbed. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the provided `###[Task Description]`, the current `###[Action]`, and prior observat... | **Turn 1 User**
###[Task Description]
Inconsistency in use `builtin.type` and `typing.Type` in error messages
Consider this example:
```py
x: type[type]
reveal_type(x) # Revealed type is "Type[builtins.type]"
reveal_type(type[type]) # Value of type "Type[type]" is not indexable
# Revealed... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "86: TypeType,\n413: if isinstance(original_type, CallableType) and original_type.is_type_obj():\n825: original_caller_type, (Instance, TupleType, TypedDictType, TypeType, CallableType)\n1215: arg_type_in_supertype_f = fo... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Act as a language world model for a web search and page-browsing environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, and prior obse... | **Turn 1 User**
###[Goal]
Which university was founded first, The Northcap University or San Diego State University? You should follow the format instruction in the requestion strictly and wrap the final answer in \boxed{}.
###[Action]
Use `browser.open` to open a search result, URL, or document section for reading.
- ... | ###[Observation]
Search results for: "NorthCap University" "established in" year
【0†Doc 12619876】 http://www.pearltrees.com/collegeindia
The NorthCap University, Sector 23a, Delhi - Itm University - Universities - Justdial. Guidelines to rate a listing Be frank and honest.
Tell us how you really feel and why. Useful r... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Function as the environment model for a mobile accessibility-control environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=306, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.google.android.apps.nexuslauncher', actions='4,8,64,16908342', child_ids='1', draw=0, win=306
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2400, pkg='com... | ###[Observation]
Android '', b=0,0,1080,2400, win=315, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.phdv.universal', actions='4,8,64,16908342', child_ids='1,2', draw=0, win=315
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.phdv.universal', actions='4,8,64... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a terminal command-execution environment. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, latest `###[Observation]`, current `###[Action]`, and earlier turns as the evidence ... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a log analyst investigating anomalous patterns in a sensor network's data. An upstream ETL job has a retry bug that produces duplicate records, and occasionally drops numerical metrics due to network jitter.
You need to write a data processing script and inte... | ###[Observation]
> > > > > > > > {"aligned_ts": 1697113814.0, "device_id": "sensor_2", "epoch_ts": 1697113814.0, "value": 50.0}
{"aligned_ts": 1697113815.0, "device_id": "sensor_2", "epoch_ts": 1697113815.0, "value": 50.0}
{"aligned_ts": 1697113815.1, "device_id": "sensor_1", "epoch_ts": 1697113815.123, "value": 10.5}
... | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Serve as a language world model for a desktop accessibility-control environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state ... | **Turn 1 User**
###[Observation]
Computer 'Task View', b=0,0,1920,1080, raw_control='WindowControl'
node=1 generic 'DesktopWindowXamlSource', raw_control='PaneControl', b=0,0,1920,1080
node=2 generic '', raw_control='PaneControl', b=0,0,1280,720
node=3 generic '', raw_control='CustomControl', automation_id='TaskV... | ###[Observation]
Computer 'Taskbar', b=0,1008,1920,1080, raw_control='PaneControl'
node=1 generic '', raw_control='PaneControl', b=1531,1008,1920,1080
node=2 generic '', raw_control='PaneControl', automation_id='40965', b=380,1008,1245,1080
node=3 generic 'Running applications', raw_control='PaneControl', b=380,100... | SFT/os/A11y-CUA | os | A11y-CUA | 3,229 | 5 | 547 | 191 | 940 | 1,678 |
Role: You are an observation generator for a language world model of a terminal command-execution environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current st... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a Database Administrator and Rust developer tasked with generating a complex performance report from an unoptimized SQLite database.
A database is located at `/home/user/company.db` containing three tables:
1. `departments` (id INTEGER PRIMARY KEY, name TEXT)... | ###[Observation]
/usr/bin/rustc
rustc 1.75.0 (82e1608df 2023-12-21) (built from a source tarball)
/usr/bin/cargo
cargo 1.75.0
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are a language world model for a software-engineering shell testbed. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the provided `###[Task Description]`, the current `###[Action]`, and a... | **Turn 1 User**
###[Task Description]
BUG: `sort_index` not preserving `index` when `ascending=False`
### Pandas version checks
- [X] I have checked that this issue has not already been reported.
- [X] I have confirmed this bug exists on the [latest version](https://pandas.pydata.org/docs/whatsnew/index.html) of pand... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "5006: def sort_index(\n5022: def sort_index(\n5038: def sort_index(\n5053: def sort_index(\n5110: def reindex(\n4902: DataFrame.sort_index : Sort a DataFrame by the index.\n5006: def sort_index(\n5022: def sort_index(\n5038:... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a mobile accessibility-control environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do not add, ski... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=882, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.duolingo', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=882
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.duolingo', actions='... | ###[Observation]
Android '', b=0,0,1080,2400, win=882, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.duolingo', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=882
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.duolingo', actions='4,8,64,16908342'... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are the simulator side of a language world model for a software-engineering editor and shell environment. The agent has already selected an action, and your job is to predict the environment response.
Input: Treat the goal or metadata, latest `###[Observation]`, current `###[Action]`, and earlier turns as th... | **Turn 1 User**
###[Goal]
<uploaded_files>
/workspace/iterative__dvc__2.8
</uploaded_files>
I've uploaded a python code repository in the directory iterative__dvc__2.8. Consider the following PR description:
<pr_description>
Negative numbers from a Python file are not recognized as parameters
# Bug Report
## dvc re... | ###[Observation]
EXECUTION RESULT of [str_replace_editor]:
Here's the result of running `cat -n` on /workspace/iterative__dvc__2.8/dvc/dependency/param.py:
1 import logging
2 import os
3 from collections import defaultdict
4
5 import dpath.util
6 from voluptuous import Any
7
8 ... | SFT/swe/swe__openhands_sft | swe | swe__openhands_sft | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a web search and page-browsing environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do not add, ski... | **Turn 1 User**
###[Goal]
Give the full name of the Spanish land grant that, in the late eighteenth century, was given to a soldier who later bequeathed it to his children. This grant covered territory west of a seasonal Arroyo in what is now the most populous city of the San Gabriel Valley, and nearly a century later ... | ###[Observation]
Search results for: "arroyo" "rancho" "west" boundary Spanish grant California 1790s
【0†Doc 5365252】 http://www.kcet.org/socal/departures/columns/maps/today-in-history-treaty-of-guadalupe-hidalgo.html
Interpretive Map of Ranchos and Adobes based on a 1937 map (see below). Unmarked areas were labeled a... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a deterministic language world model for a mobile accessibility-control environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply t... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=833, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.ebay.mobile', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=833
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.ebay.mobile', act... | ###[Observation]
Android '', b=0,0,1080,2400, win=833, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.ebay.mobile', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=833
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.ebay.mobile', actions='4,8,64,169... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: Serve as a language world model for an MCP business-tool environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state snapshot. A... | **Turn 1 User**
###[Goal]
Fetch all orders where I am the seller that are currently in 'awaiting_shipment' status, create shipment records for each using carrier 'UPS', generate dummy tracking numbers 'UPS-TEST-1001', 'UPS-TEST-1002', etc., set initial delivery status to 'label_created', and update the corresponding or... | ###[Observation]
{
"id": "ORD-LAP-1001",
"lot_id": "LOT-LAP-1001",
"buyer_id": 1,
"seller_id": 1,
"final_price": 520.0,
"status": "shipped",
"created_at": "2026-06-21T12:07:36",
"updated_at": "2026-07-03T12:07:59.112964"
} | SFT/mcp/AWM | mcp | AWM | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a desktop accessibility-control environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current... | **Turn 1 User**
###[Observation]
Computer 'Documents - File Explorer', b=370,124,1570,1024, raw_control='WindowControl'
node=1 generic '', raw_control='PaneControl', b=812,124,1559,175
node=2 generic 'Documents', raw_control='PaneControl', b=381,328,1559,1013
node=3 generic '', raw_control='PaneControl', b=381,328,... | ###[Observation]
Computer 'Sign in | Slack', b=114,0,1806,1128, raw_control='PaneControl'
node=1 document 'Sign in | Slack', raw_control='DocumentControl', automation_id='585200', b=114,0,1806,1128
node=2 group '', raw_control='GroupControl', b=114,0,1806,1175
node=3 group '', raw_control='GroupControl', b=114,0,... | SFT/os/A11y-CUA | os | A11y-CUA | 9,106 | 0 | 491 | 1,128 | 342 | 1,961 |
Role: Serve as a language world model for a software-engineering shell testbed. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state snapsho... | **Turn 1 User**
###[Task Description]
Dask interprets boolean array as an index array
**What happened**: setting items on a dask array with a bool array incorrectly casts the array to an index array.
**What you expected to happen**: no broadcastig error.
**Minimal Complete Verifiable Example**:
```python
>>> im... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "numpy mask OK\n", "stderr": "", "duration_ms": 996} | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a language world model for a mobile accessibility-control environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=1818, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.rightmove.android', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1818
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.rightmove... | ###[Observation]
Android '', b=0,0,1080,2400, win=1818, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.rightmove.android', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1818
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.rightmove.android', actio... | SFT/android/AndroidControl | android | AndroidControl | null | null | null | null | null | null |
Role: You are an observation generator for a language world model of a software-engineering shell testbed. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the current state.... | **Turn 1 User**
###[Task Description]
BUG: PeriodIndex does not survive JSON serialization roundtrip
###
- [X] I have checked that this issue has not already been reported.
- [X] I have confirmed this bug exists on the [latest version](https://pandas.pydata.org/docs/whatsnew/index.html) of pandas.
- [x] I hav... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": " df = df.set_index(table[\"schema\"][\"primaryKey\"])\n if len(df.index.names) == 1:\n if df.index.name == \"index\":\n df.index.name = None\n else:\n df.index.names = [\n None if... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: You are a post-action environment simulator for a terminal command-execution environment. The input action is already fixed; your role is to predict what the environment exposes next.
Input: Use the available goal, metadata, latest `###[Observation]`, action text, and history as state evidence. Do not add, skip,... | **Turn 1 User**
###[Goal]
Please solve this task:
You are an automated backup operator assistant. We need to implement a continuous restore-testing service in user-space to automatically verify incoming backup files.
Since we do not have root access or a standard `systemd` environment, you will build a custom Python... | ###[Observation]
Service started
Service is running (PID: 61)
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: You are a language world model for a web search and page-browsing environment. Your responsibility is to simulate the environment after an action has already been issued, not to act as the agent.
Input: This prompt accepts the goal or task context, the latest `###[Observation]`, the current `###[Action]`, and an... | **Turn 1 User**
###[Goal]
A doctoral thesis addressing mechanisms behind Earth's self-regulation was published open access in 2019 by a UK university within the world's most relied-upon graduate works repository. As of 2023, approximately how many libraries worldwide use this resource provider, and how many completely ... | ###[Observation]
Search results for: "self-regulation" "Earth" PhD thesis 2019 open access
【0†Doc 7576780】 http://sens-de-la-vie.com/Archepages/vsv_relationship_human-ea.htm
Les idées contenues dans le site prennent vie !
Après 12 ans d’existence, près de 20 millions de visites et un travail colossal réalisé par son a... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: You are a deterministic language world model for a terminal command-execution environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply the... | **Turn 1 User**
###[Goal]
Please solve this task:
You are acting as a network security engineer investigating a recent security incident. We captured some suspicious network traffic on our internal network, and we suspect a local web application has been compromised.
Your tasks are as follows:
1. **Traffic Analysis*... | ###[Observation]
> > > > > > > > > > > > > > > > > > > > > > > > Done searching
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Act as a language world model for a terminal command-execution environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, and prior observ... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a data engineer working on an ETL pipeline for a climate monitoring station. We receive noisy, irregular temperature readings and need to clean, aggregate, and resample them into a regular time series.
You have been provided with a vendored C utility located a... | ###[Observation]
total 0
drwxr-xr-x 2 root root 60 May 11 07:40 .
drwxr-xr-x 3 root root 80 May 11 07:40 ..
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Serve as a language world model for a terminal command-execution environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state sna... | **Turn 1 User**
###[Goal]
Please solve this task:
You are a database administrator working on optimizing recursive queries for an older relational database. The database stores a hierarchical graph (a forest of trees representing an organizational structure). To speed up read-heavy queries, you need to write a Rust ut... | ###[Observation]
Created binary (application) `graph_optimizer` package
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Operate as a language world model for a terminal command-execution environment. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. The ... | **Turn 1 User**
###[Goal]
Please solve this task:
A researcher in our lab is running large-scale finite element simulations. We use a domain decomposition method with adaptive mesh refinement. However, the simulation often crashes because certain mesh subdomains produce near-singular matrices during the matrix factori... | ###[Observation]
1.262726 2.071514
1.374782 2.056401
1.576944 1.848753
1.257289 1.700323
0.905860 1.695319
(exit_code=0) | SFT/terminal/TMax | terminal | TMax | null | null | null | null | null | null |
Role: Operate as a language world model for a software-engineering shell testbed. You model how the environment changes after an issued action and report only the resulting observation.
Input: Interpret the goal or task metadata, latest observation, action block, and earlier turns as a persistent state trace. The curr... | **Turn 1 User**
###[Task Description]
[bug] `conan profile show` crashes when requesting json output if profile contains a reference.
### Environment details
* Operating System+version: macOS 12.6
* Conan version: 2.0.14
* Python version: 3.11.4
### Steps to reproduce
Issuing `conan profile show -pr:h ourprofile... | ###[Observation]
{"status": "SUCCEEDED", "exit_code": 0, "stdout": "1:import json\n38: tc.run(\"profile show -h\")\n44: \"profile show -o:a=both_options=True -pr:a=profile -s:a=os=WindowsCE -s:a=os.platform=conan -c:a=user.conf.cli=True -f=json\")\n46: out = json.loads(tc.out)\n52: ... | SFT/swe/SWE-Gym | swe | SWE-Gym | null | null | null | null | null | null |
Role: Function as the environment model for a web search and page-browsing environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the... | **Turn 1 User**
###[Goal]
Identify the father of Anna De' Medici, Archduchess Of Austria, and then determine the name of her paternal grandmother.
###[Action]
Use `browser.open` to open a search result, URL, or document section for reading.
- This advances the task or changes state.
- Args: `id`="https://en.wikipedia.o... | ###[Observation]
Search results for: dbpedia "Anna de' Medici"
【0†Doc 11534678】 https://www.futuretools.io/tools/perplexity-ai
A new search interface that uses OpenAI GPT 3.5 and Microsoft Bing to directly answer any question you ask.
🚩 WARNING: This tool has been flagged for either trying to game the upvote system, ... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Act as a language world model for a metadata-only MCP analytics and notebook environment. You predict the post-action observation of the environment; you are not the acting agent, planner, or tool executor.
Input: Use the goal or task description, the most recent `###[Observation]`, the current `###[Action]`, an... | **Turn 1 User**
###[Observation]
{"key": "task_00149_comp_pos_079200f5", "env_key": "carlisle"}
###[Action]
Use `list_workspace_files` to retrieve or inspect list workspace files information.
- This advances the task or changes state.
- Args: `pattern`="*.ipynb".
- Call:
```text
list_workspace_files(pattern="*.ipynb")
... | ###[Observation]
{"type": "text", "text": "{\n \"ok\": true,\n \"text\": \"Cell 13 (code) deleted successfully.\\ndeleted cell source:\\nt_stat_orig, p_value_orig = stats.ttest_ind(\\n df_baseline['ltv_90d_usd'],\\n df_interest['ltv_90d_usd']\\n)\\nprint(f\\\"Original p-value: {p_value_orig:.4f}\\\")\\n\\n---\\... | SFT/mcp/mcp__fleet_tool_use_difficult_envs | mcp | mcp__fleet_tool_use_difficult_envs | null | null | null | null | null | null |
Role: You are a deterministic language world model for a mobile accessibility-control environment. Predict the environment observation caused by the already-issued action, and nothing else.
Input: Use all provided goal or task context, recent observation, current action, and previous turns as the state record. Apply t... | **Turn 1 User**
###[Observation]
Android '', b=0,0,1080,2400, win=1391, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=0,0,1080,2400, pkg='com.andrwq.recorder', actions='4,8,64,16908342', child_ids='1,2,3', draw=0, win=1391
node=2 group '', raw_class='LinearLayout', b=0,0,1080,2337, pkg='com.andrwq.reco... | ###[Observation]
Android '', b=28,656,1052,1799, win=1392, layer=0, type=1
node=1 group '', raw_class='FrameLayout', b=28,656,1052,1799, pkg='com.andrwq.recorder', actions='4,8,64,16908342', child_ids='1', draw=0, win=1392
node=2 group '', raw_class='LinearLayout', b=70,698,1010,1757, pkg='com.andrwq.recorder', acti... | SFT/android/AndroidControl | android | AndroidControl | 10,630 | 4 | 507 | 2,600 | 1,156 | 4,263 |
Role: Function as the environment model for a browser accessibility environment with reference IDs. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction sta... | **Turn 1 User**
###[Observation]
- RootWebArea "Today: Todoist" [ref=e1]
- link "Skip to task list" [ref=e2]
- banner "Top Banner: contains Search, Quick Add, Productivity, Help, Notifications, and Settings" [ref=e3]
- switch "Main menu" [ref=e4]
- graphics-symbol [ref=e5]
- button "Go to Home view" [... | ###[Observation]
- RootWebArea "Today: Todoist" [ref=e1]
- link "Skip to task list" [ref=e2]
- switch "Main menu" [ref=e3]
- button "Go to Home view" [ref=e4]
- combobox "Quick search" [ref=e5]
- link "How to use search" [ref=e6]
- button "Close Search" [ref=e7]
- button "Quick search" [ref=e8]
- button... | SFT/web/web__weblinx_browsergym | web | web__weblinx_browsergym | 14,988 | 0 | 608 | 2,900 | 1,171 | 4,679 |
Role: You are an observation generator for a language world model of a software-engineering editor and shell environment. The action has already occurred; simulate the environment result rather than agent intent.
Input: Use the goal or task context, current observation, action block, and prior turns to reconstruct the... | **Turn 1 User**
###[Goal]
<uploaded_files>
/workspace/getmoto__moto__4.2
</uploaded_files>
I've uploaded a python code repository in the directory getmoto__moto__4.2. Consider the following PR description:
<pr_description>
stepfunctions list_executions with aborted execution does not return stopDate attribute in respo... | ###[Observation]
EXECUTION RESULT of [str_replace_editor]:
The file /workspace/getmoto__moto__4.2/moto/stepfunctions/responses.py has been edited. Here's the result of running `cat -n` on a snippet of /workspace/getmoto__moto__4.2/moto/stepfunctions/responses.py:
144 "stateMachineArn": state_machine.... | SFT/swe/swe__openhands_sft | swe | swe__openhands_sft | null | null | null | null | null | null |
Role: Function as the environment model for a web search and page-browsing environment. Given the action already chosen by another agent, infer the next visible observation without taking agency yourself.
Input: Read the goal or metadata, current observation, action, and history as a single interaction state. Only the... | **Turn 1 User**
###[Goal]
A prominent leader of a Salafi jihadist group, noted for using elaborate video concealment techniques around five years after the early 2010s death of his predecessor, succeeded the founder who was killed in early May 2011 during a special operation in a northern Pakistani city. This city was ... | ###[Observation]
Search results for: Abbottabad history dawn.com "James Abbott"
【0†Doc 5018343】 http://www.genealogy.com/forum/surnames/topics/abbott/3904/
Hi Cat. It is a moot subject now, However, my question was based on a statement from a book by Othman A Abbott titled "Recolletions 0f a Pioneer Lawer" reprinted i... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Role: Serve as a language world model for a web search and page-browsing environment. Your output should be the environment state immediately after the recorded action, not the agent behavior before or after it.
Input: Combine the goal or task context, current observation, issued action, and prior turns into a state s... | **Turn 1 User**
###[Goal]
A French battalion that fought unsuccessfully in a major southern U.S. city in 1779 later sailed to a region administered at various times by the British, French, Spanish, Dutch, and Danish, where it participated in further campaigns. This region's name was adopted by several powerful European... | ###[Observation]
Search results for: "Windward Passage" pirate ship seized 18th century Cuba Hispaniola
【0†Doc 954180】 https://www.newspapers.com/clip/109598/harrisburg_telegraph_6november_1873/
Harrisburg Telegraph 6.November 1873
FROM HAVANA. CAPTURE OF THE YIRGINIUS. ALL ON BOARD MADE PRISONERS. TO BE TRIED FOR PIR... | SFT/search/OpenResearcher | search | OpenResearcher | null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.