johannhartmann/decider-2b-a11y-crawler-lora
Updated • 46 • 1
task_id stringlengths 9 12 | split stringclasses 2
values | objective stringlengths 3 2.78k | start_url stringclasses 18
values | end_url stringlengths 1 5.42k | n_steps int64 1 37 | decision stringclasses 2
values | objective_coherent bool 2
classes | trajectory_consistent bool 2
classes | reason stringlengths 83 403 | model stringclasses 1
value | prompt_trajectory stringlengths 73 5.77k |
|---|---|---|---|---|---|---|---|---|---|---|---|
openweb_0 | train | Find a chicken recipe. | https://www.google.com/ | https://littlesunnykitchen.com/marry-me-chicken/ | 3 | keep | true | true | Clear objective to find a chicken recipe. Trajectory searches for chicken recipes and lands on a specific chicken recipe page, matching the goal. | claude-haiku-4-5-20251001 | page: https://www.google.com/
1. type [95] [chicken recipe ] where [95] is Search
page: https://www.google.com/search?q=chicken+recipe&sca_esv=53e635cf3415af69&source=hp&ei=vkl7Z9DgCejJkPIPoPzE8A0&iflsig=AL9…
2. click [569] where [569] is Marry Me Chicken Little Sunny Kitchen Rated 5.0 out of 5, (2.1K) 25 min Chi... |
openweb_1 | train | Look for chicken recipes. | https://www.google.com/ | https://www.cookingclassy.com/?s=chicken | 9 | keep | true | true | Objective is clear and coherent. Trajectory successfully searches for chicken recipes on Google, navigates to Cooking Classy, and performs searches specifically for chicken recipes, ending at results page. | claude-haiku-4-5-20251001 | page: https://www.google.com/
1. type [112] [cooking recipes ] where [112] is Search
page: https://www.google.com/search?q=cooking+recipes&sca_esv=c8f39132a967e6fa&source=hp&ei=hGhvZ4XhDK7PkPIPlrr4qAM&iflsig=AL…
2. click [541] where [541] is Cooking Classy: Delicious, Easy-to-Make Recipes ... Cooking Classy https... |
openweb_10 | train | Search for something on the BBC News website. | https://www.bbc.com/news | https://www.bbc.com/search?q=latest%20sports%20news&edgeauth=eyJhbGciOiAiSFMyNTYiLCAidHlwIjogIkpXVCJ9.eyJrZXkiOiAiZmFzdGx5LXVyaS10b2tlbi0xIiwiZXhwIjogMTczNjIyMjYzNiwibmJmIjogMTczNjIyMjI3NiwicmVxdWVzdHVyaSI6ICIlMkZzZWFyY2glM0ZxJTNEbGF0ZXN0JTI1MjBzcG9ydHMlMjUyMG5ld3MifQ.G3SbI_vgCUcnWxuSMQWdAiMDB9caTvzsReM73IwfiJo | 20 | keep | true | true | Clear objective to search BBC News. Trajectory executes multiple searches, eventually landing on 'latest sports news' search results matching the objective. | claude-haiku-4-5-20251001 | page: https://www.bbc.com/news
1. click [125] where [125] is Search BBC
2. type [235] [news ] where [235] is Search news, topics and more
page: https://www.bbc.com/search?q=news&edgeauth=eyJhbGciOiAiSFMyNTYiLCAidHlwIjogIkpXVCJ9.eyJrZXkiOiAiZmFzdGx5LXVyaS10b2tlbi0…
3. press [Enter]
4. click [246] where [246] is Cl... |
openweb_100 | train | Find out how to upgrade from the free version of GitHub Copilot to the pro version. | https://github.com/ | https://github.com/login?return_to=https%3A%2F%2Fgithub.com%2Fsettings%2Fcopilot%3Futm_campaign%3Dcopilotfree-bannerheader%26utm_medium%3Dbanner%26utm_source%3Dgithub | 13 | keep | true | true | Objective is clear and coherent. Trajectory directly pursues learning about upgrading from free to pro Copilot, including checking FAQs and attempting to access settings. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [156] where [156] is Learn more
page: https://github.com/features/copilot/?utm_source=github&utm_medium=banner&utm_campaign=copilotfree-bannerheader
2. hover [660] where [660] is
3. click [372] where [372] is
page: https://github.com/pricing?utm_campaign=copilotfree-bannerhea... |
openweb_1000 | train | Find the results of the research on Gamma-Ray Burst Radiation Mechanisms from the article titled "New Insights on Gamma-Ray Burst Radiation Mechanisms from Multiwavelength Observations". | https://arxiv.org/ | https://arxiv.org/html/2412.18577v1#S1 | 37 | keep | true | true | Agent navigates to correct article and explores its sections (Abstract, Introduction, Results, Methods) to find research results on GRB radiation mechanisms, successfully locating relevant findings. | claude-haiku-4-5-20251001 | page: https://arxiv.org/
1. click [220] where [220] is Physics
page: https://arxiv.org/archive/physics
2. type [52] [astrophysics ] where [52] is Search term or terms
page: http://arxiv.org/search/?query=astrophysics&searchtype=all&source=header
3. click [1090] where [1090] is
page: https://arxiv.org/abs/24... |
openweb_1001 | train | Search for bike-friendly restaurants near Stanford. | https://www.google.com/maps | https://www.google.com/maps/search/bike-friendly+restaurants+near+Stanford/@37.4352519,-122.168975,15z?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 5 | keep | true | true | Objective is clear and well-formed. Trajectory directly pursues it via Google Maps search refinement, successfully reaching bike-friendly restaurants near Stanford. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [restaurants near me ] where [142] is Search Google Maps
page: https://www.google.com/maps/search/restaurants+near+me/@37.4767616,-122.2017024,12z?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXM…
2. type [142] [restaurants near Stanford ] where [142] is
page: https://www.goo... |
openweb_1002 | train | Find highly-rated bike-friendly restaurants near Stanford. | https://www.google.com/maps | https://www.google.com/maps/search/bike-friendly+restaurants+near+Stanford/@37.4352492,-122.1689751,15z/data=!3m1!4b1!4m2!2m1!6e5?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 9 | keep | true | true | Agent searches for bike-friendly restaurants near Stanford, filters by rating, and reaches results page. Trajectory directly pursues the stated objective. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [restaurants near me ] where [142] is Search Google Maps
page: https://www.google.com/maps/search/restaurants+near+me/@37.4767616,-122.2017024,12z?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXM…
2. type [142] [restaurants near Stanford ] where [142] is
page: https://www.goo... |
openweb_1003 | train | Find "The Melt" restaurant near Stanford. | https://www.google.com/maps | https://www.google.com/maps/place/The+Melt/@37.4333076,-122.1853184,14z/data=!4m6!3m5!1s0x808fbb343b58d50f:0x959033ad7d29ceba!8m2!3d37.4430947!4d-122.1725973!16s%2Fg%2F1hc243ntl?entry=ttu&g_ep=EgoyMDI1MDEwMS4wIKXMDSoASAFQAw%3D%3D | 2 | keep | true | true | Objective is clear and well-formed. Trajectory directly searches for and finds The Melt restaurant near Stanford on Google Maps as requested. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [The Melt near Stanford ] where [142] is Search Google Maps
page: https://www.google.com/maps/place/The+Melt/@37.4333076,-122.1853184,14z/data=!4m6!3m5!1s0x808fbb343b58d50f:0x959033ad7d…
2. stop [Found the information about The Melt restaurant near Stanford on the G... |
openweb_1004 | heldout | Find information about adjectives in English grammar. | https://dictionary.cambridge.org/ | https://dictionary.cambridge.org/grammar/british-grammar/adjectives_1?q=Adjectives | 5 | keep | true | true | Objective is clear and well-formed. Trajectory successfully navigates to the adjectives grammar page on Cambridge Dictionary, directly pursuing the stated goal. | claude-haiku-4-5-20251001 | page: https://dictionary.cambridge.org/
1. click [170] where [170] is Dictionary
page: https://dictionary.cambridge.org/dictionary/
2. click [164] where [164] is Grammar
page: https://dictionary.cambridge.org/grammar/british-grammar/
3. type [260] [adjectives ] where [260] is Search
page: https://dictionary... |
openweb_1005 | heldout | Find out how to use adjectives in phrases. | https://dictionary.cambridge.org/ | https://dictionary.cambridge.org/grammar/british-grammar/adjective-phrases | 5 | keep | true | true | Objective clearly asks about adjective phrases. Trajectory navigates through grammar sections directly to the adjective phrases page and stops appropriately. | claude-haiku-4-5-20251001 | page: https://dictionary.cambridge.org/
1. click [173] where [173] is Grammar
page: https://dictionary.cambridge.org/grammar/british-grammar/
2. click [301] where [301] is Adjectives and adverbs
page: https://dictionary.cambridge.org/grammar/british-grammar/adjectives-and-adverbs
3. click [300] where [300] is ... |
openweb_1007 | train | Find the profile of Scottie Barnes from the Toronto Raptors. | https://www.espn.com/ | https://www.espn.com/nba/player/_/id/4433134/scottie-barnes | 9 | keep | true | true | Objective is clear and coherent. Trajectory successfully navigates to Scottie Barnes' ESPN player profile, achieving the stated goal. | claude-haiku-4-5-20251001 | page: https://www.espn.com/
1. click [535] where [535] is NBA
2. click [6742] where [6742] is Full Scoreboard »
page: https://www.espn.com/nba/scoreboard
3. click [348] where [348] is Full Scoreboard »
4. scroll [down]
5. click [1680] where [1680] is
6. click [865] where [865] is
page: https://www.espn.com/nba... |
openweb_1008 | train | Find Toronto Raptors team information. | https://www.espn.com/ | https://www.espn.com/nba/team/_/name/tor/toronto-raptors | 4 | keep | true | true | Objective is clear and well-formed. Trajectory directly searches for and navigates to Toronto Raptors team information page on ESPN. | claude-haiku-4-5-20251001 | page: https://www.espn.com/
1. click [1830] where [1830] is Open Search
2. type [1833] [Toronto Raptors ] where [1833] is Search
page: https://www.espn.com/search/_/q/Toronto%20Raptors
3. click [289] where [289] is Toronto Raptors Toronto Raptors NBA
page: https://www.espn.com/nba/team/_/name/tor/toronto-rapto... |
openweb_1009 | train | Ask Hugging Face how many datasets support Qwen models. | https://huggingface.co/ | https://huggingface.co/datasets?sort=trending&search=Qwen+models | 7 | keep | true | true | Agent searches for datasets related to Qwen models and reaches a filtered results page. Trajectory pursues the objective despite some navigation detours. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. type [44] [Qwen ] where [44] is Search models, datasets, users...
page: https://huggingface.co/Qwen/QVQ-72B-Preview
2. click [152] where [152] is Datasets
page: https://huggingface.co/datasets
3. click [242] where [242] is Full-text search
page: https://huggingface.co/search... |
openweb_1010 | train | Find out more about the Qwen model's capabilities and limitations. | https://huggingface.co/ | https://huggingface.co/Qwen/QwQ-32B-Preview/blob/main/README.md?code=true#L16 | 15 | keep | true | true | The objective is clear: find information about Qwen model capabilities and limitations. The trajectory searches for Qwen models on Hugging Face and navigates to relevant model cards that contain this information. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. type [44] [Qwen model ] where [44] is Search models, datasets, users...
page: https://huggingface.co/deepseek-ai/DeepSeek-V3
2. click [149] where [149] is Models
page: https://huggingface.co/models
3. click [878] where [878] is Qwen/Qwen2-VL-7B-Instruct
page: https://hugging... |
openweb_1011 | train | Find out what the Qwen model can do. | https://huggingface.co/ | https://huggingface.co/Qwen/QVQ-72B-Preview | 2 | keep | true | true | User searches for Qwen model information, reaches its HuggingFace page, and reads about its capabilities. Coherent objective with matching trajectory. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. type [44] [Qwen ] where [44] is Search models, datasets, users...
page: https://huggingface.co/Qwen/QVQ-72B-Preview
2. stop [The Qwen model is mainly for visual reasoning tasks, including but not limited to: enhancing visual reasoning capabilities; multidisciplinary understanding ... |
openweb_1012 | train | What are the capabilities of the Qwen/QwQ-32B-Preview model (specifically, can it translate from languages like Spanish, French, or Chinese)? | https://huggingface.co/ | https://huggingface.co/Qwen/QwQ-32B-Preview | 7 | drop | true | false | Trajectory clicks 'Datasets' then searches for the model, but never actually interacts with the model's capabilities documentation or interface to answer the question asked. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. click [55] where [55] is Datasets
page: https://huggingface.co/datasets
2. type [76] [Qwen/QwQ-32B-Preview ] where [76] is Search models, datasets, users...
page: https://huggingface.co/Qwen/QwQ-32B-Preview
3. type [479] [Can you translate text from languages such as Spanish, F... |
openweb_1013 | train | Find math college courses online. | https://www.coursera.org/ | https://www.coursera.org/search?query=math%20college%20courses | 3 | keep | true | true | The objective is clear and well-formed. The trajectory plausibly pursues it by searching for math college courses on Coursera and clicking a relevant result. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2057] [math college courses ] where [2057] is What do you want to learn?
page: https://www.coursera.org/search?query=math%20college%20courses
2. click [799] where [799] is Math Prep: College & Work Ready free Course by University of North Texas, 4.5 stars, by 251 reviews, ... |
openweb_1015 | train | Find out the course content for Data Science Math Skills on Coursera. | https://www.coursera.org/ | https://www.coursera.org/learn/datasciencemathskills#modules | 13 | keep | true | true | Demonstrator successfully navigates to Data Science Math Skills course and accesses the modules section, viewing course content structure despite access limitations. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2058] [Math ] where [2058] is What do you want to learn?
page: https://www.coursera.org/search?query=Math
2. click [753] where [753] is Math Prep: College & Work Ready free Course by University of North Texas, 4.5 stars, by 250 reviews, provide skills Algebra, Mathema…
... |
openweb_1016 | train | Find the instructor for the "Data Science Math Skills" course. | https://www.coursera.org/ | https://www.coursera.org/learn/datasciencemathskills | 3 | keep | true | true | Agent searched for the course, navigated to its page, and identified the instructors Daniel Egger and Paul Bendich, directly fulfilling the objective. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2058] [Data Science Math Skills ] where [2058] is What do you want to learn?
page: https://www.coursera.org/search?query=Data%20Science%20Math%20Skills
2. click [800] where [800] is Data Science Math Skills
page: https://www.coursera.org/learn/datasciencemathskills
3. s... |
openweb_1017 | train | Find out what courses are available for someone interested in 'Data Science Math Skills' at Duke University. | https://www.coursera.org/ | https://www.coursera.org/learn/datasciencemathskills | 6 | keep | true | true | Objective is clear: find Data Science Math Skills courses at Duke. Trajectory searches for and locates the course, explores Duke's offerings, demonstrating relevant exploration. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2058] [Data Science Math Skills at Duke University ] where [2058] is What do you want to learn?
page: https://www.coursera.org/search?query=Data%20Science%20Math%20Skills%20at%20Duke%20University
2. click [799] where [799] is Data Science Math Skills free Course by Duke Un... |
openweb_1019 | train | Find out more about the Data Science Math Skills course at Coursera and the admission process for its related field of study at Duke University. | https://www.coursera.org/ | https://admissions.duke.edu/apply/#requirements | 29 | keep | true | true | Agent explores Data Science Math Skills course on Coursera, then navigates to Duke admissions application page showing admission requirements as instructed. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2058] [Math ] where [2058] is What do you want to learn?
page: https://www.coursera.org/search?query=Math
2. click [753] where [753] is Math Prep: College & Work Ready free Course by University of North Texas, 4.5 stars, by 250 reviews, provide skills Algebra, Mathema…
... |
openweb_1021 | train | Research and gather information about the application process and deadlines for Duke University's undergraduate admissions. | https://www.coursera.org/ | https://admissions.duke.edu/apply/ | 4 | keep | true | true | Objective is clear and well-formed. Trajectory navigates to Duke admissions site and reaches the apply page with deadline information, directly pursuing the stated research goal. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. new_tab
page: about:blank
2. goto [https://admissions.duke.edu]
page: https://admissions.duke.edu/
3. click [519] where [519] is LEARN MORE ABOUT ADMISSIONS Arrow Right
page: https://admissions.duke.edu/apply/
4. stop [The webpage contains information regarding deadlines: ... |
openweb_1023 | train | What is myBMI if I am 5 ft 9 in and I weigh 170 lbs? | https://www.wolframalpha.com/ | https://www.wolframalpha.com/input?i=BMI&assumption=%7B%22F%22%2C+%22BodyMassIndex%22%2C+%22H%22%7D+-%3E%225+ft+9+in%22&assumption=%7B%22FS%22%7D+-%3E+%7B%7B%22BodyMassIndex%22%2C+%22BMI%22%7D%2C+%7B%22BodyMassIndex%22%2C+%22H%22%7D%2C+%7B%22BodyMassIndex%22%2C+%22W%22%7D%7D&assumption=%7B%22F%22%2C+%22BodyMassIndex%22... | 9 | keep | true | true | User asks to calculate BMI given height and weight. Trajectory uses WolframAlpha to input these values and retrieve the BMI result. | claude-haiku-4-5-20251001 | page: https://www.wolframalpha.com/
1. type [89] [BMI ] where [89] is WolframAlpha input field
page: https://www.wolframalpha.com/input?i=BMI
2. type [508] [170 lb] where [508] is
3. type [512] [5 ft 9 in ] where [512] is
page: https://www.wolframalpha.com/input?i=BMI&assumption=%7B%22F%22%2C+%22BodyMassIndex%... |
openweb_1027 | train | Calculate my BMI. | https://www.wolframalpha.com/ | https://www.wolframalpha.com/input?i=bmi+calculator&assumption=%7B%22F%22%2C+%22BodyMassIndex%22%2C+%22H%22%7D+-%3E%225+ft+8+in%22&assumption=%7B%22FS%22%7D+-%3E+%7B%7B%22BodyMassIndex%22%2C+%22BMI%22%7D%2C+%7B%22BodyMassIndex%22%2C+%22H%22%7D%2C+%7B%22BodyMassIndex%22%2C+%22W%22%7D%7D&assumption=%7B%22F%22%2C+%22BodyM... | 5 | keep | true | true | Objective is clear and coherent. Trajectory uses WolframAlpha's BMI calculator with height and weight inputs, reaching a calculated BMI result of 23.42. | claude-haiku-4-5-20251001 | page: https://www.wolframalpha.com/
1. click [96] where [96] is NATURAL LANGUAGE
2. type [89] [BMI calculator ] where [89] is WolframAlpha input field
page: https://www.wolframalpha.com/input?i=bmi+calculator
3. type [493] [5 ft 8 in] where [493] is
4. type [489] [154 lb ] where [489] is
page: https://www.wolf... |
openweb_1028 | train | Calculate the BMI for a person who is 5'9" and weighs 170 lbs, and find more information about this calculated value. | https://www.wolframalpha.com/ | about:blank | 29 | keep | true | true | Agent successfully calculated BMI (25.1) for given measurements and explored related information on WolframAlpha before navigation errors occurred. | claude-haiku-4-5-20251001 | page: https://www.wolframalpha.com/
1. type [89] [BMI ] where [89] is WolframAlpha input field
page: https://www.wolframalpha.com/input?i=BMI
2. type [508] [170 lb] where [508] is
3. type [512] [5 ft 9 in ] where [512] is
page: https://www.wolframalpha.com/input?i=BMI&assumption=%7B%22F%22%2C+%22BodyMassIndex%... |
openweb_1029 | train | Find the body mass index (BMI) of a person who is 5 ft 9 in (170 lbs) and analyze the result. | https://www.wolframalpha.com/ | https://www.wolframalpha.com/input?i=BMI+5+ft+9+in+170+lbs | 2 | keep | true | true | Objective asks to find and analyze BMI for given measurements. Trajectory successfully queries WolframAlpha with correct values and arrives at result showing overweight classification. | claude-haiku-4-5-20251001 | page: https://www.wolframalpha.com/
1. type [89] [BMI 5 ft 9 in 170 lbs ] where [89] is WolframAlpha input field
page: https://www.wolframalpha.com/input?i=BMI+5+ft+9+in+170+lbs
2. stop [overweight] (final recorded decision) |
openweb_103 | train | Find a specific arXiv paper on machine learning and access its different formats. | https://arxiv.org/ | https://arxiv.org/abs/2412.18601 | 5 | keep | true | true | Agent searches for ML paper, finds arXiv:2412.18601, navigates to it, and explores format options. Actions align with objective. | claude-haiku-4-5-20251001 | page: https://arxiv.org/
1. type [50] [machine learning ] where [50] is Search term or terms
page: http://arxiv.org/search/?query=machine+learning&searchtype=all&source=header
2. click [205] where [205] is arXiv:2412.18601
page: https://arxiv.org/abs/2412.18601
3. click [219] where [219] is
4. click [276] wher... |
openweb_1030 | train | Search for flights on January 1, 2025, to January 3, 2025. | https://www.google.com/travel/flights | https://www.google.com/travel/flights?tfs=CBwQARoaEgoyMDI1LTAxLTAxagwIAhIIL20vMGQ2bHAaGhIKMjAyNS0wMS0wM3IMCAISCC9tLzBkNmxwQAFIAXABggELCP___________wGYAQE&tfu=KgIIAw | 5 | keep | true | true | Objective is clear and well-formed. Trajectory enters correct departure and return dates on Google Flights matching the objective dates, reaching expected results page. | claude-haiku-4-5-20251001 | page: https://www.google.com/travel/flights
1. type [691] [2025-01-01 ] where [691] is Departure
page: https://www.google.com/travel/flights?tfs=CBwQARoaEgoyMDI1LTAxLTAxagwIAhIIL20vMGQ2bHAaDnIMCAISCC9tLzBkNmxwQAFIAXABggELC…
2. type [712] [2025-01-03 ] where [712] is Return
page: https://www.google.com/travel/f... |
openweb_1033 | train | Find out how Transformers work in the Hugging Face NLP course. | https://huggingface.co/ | https://huggingface.co/learn/nlp-course/chapter1/5?fw=pt | 9 | keep | true | true | Objective clearly asks to learn how Transformers work in HF NLP course. Trajectory navigates through course chapters, reaching detailed Transformer content as requested. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. click [615] where [615] is Learn
page: https://huggingface.co/learn
2. click [82] where [82] is NLP Course This course will teach you about natural language processing using libraries from the HF ecosystem.
page: https://huggingface.co/learn/nlp-course/chapter1/1
3. click [185]... |
openweb_1035 | heldout | Search for "Winter Sale toys". | https://www.amazon.com/ | https://www.amazon.com/s?k=Winter+Sale+toys&crid=VBT80R7R6DQA&sprefix=winter+sale+toys%2Caps%2C164&ref=nb_sb_noss_1 | 3 | keep | true | true | Objective is clear and coherent. Trajectory searches for 'Winter Sale toys' on Amazon, reaching the search results page as intended. The subsequent filter action is a normal refinement. | claude-haiku-4-5-20251001 | page: https://www.amazon.com/
1. type [239] [Winter Sale toys ] where [239] is Search Amazon
page: https://www.amazon.com/s?k=Winter+Sale+toys&crid=VBT80R7R6DQA&sprefix=winter+sale+toys%2Caps%2C164&ref=nb_sb_noss_1
2. type [152] [Toys & Games ] where [152] is
3. stop [] (final recorded decision) |
openweb_1036 | heldout | Find the cheapest stuffed animals in the Winter Sale. | https://www.amazon.com/ | https://www.amazon.com/Christmas-Keychain-Birthday-Decoration-White-red/dp/B0DHTQ8V7P/ref=sr_1_1?crid=3DHRPG43W6TVE&dib=eyJ2IjoiMSJ9.zgUu-fffZ-VdnZCzHTLN0D8L2hE0LWAxjZlWzymLD8L1P2pc1l5cG1unxn4oxKpEwgSDN-V7AtDbhqMzzpmsmAK92dAelDvEU1KhMYNzRDWt7QKCb7pnFlFbfZnhauii4Ahqc4MJG1eIvNAJff2VT007fj4h_Sl7EOLMhnB-SkB73z2gm9aotoYk6EF... | 9 | drop | true | false | Trajectory stops at single product without comparing prices to find cheapest. Goal requires price comparison; trajectory doesn't achieve it. | claude-haiku-4-5-20251001 | page: https://www.amazon.com/
1. click [184] where [184] is Search Amazon
2. type [184] [Winter Sale toys ] where [184] is Search Amazon
page: https://www.amazon.com/s?k=Winter+Sale+toys&crid=15FG1S7D9BO87&sprefix=winter+sale+toys%2Caps%2C131&ref=nb_sb_noss
3. hover [86] where [86] is
4. type [86] [Toy Deals] whe... |
openweb_1037 | heldout | Find and purchase low-cost stuffed animals from the Winter Sale. | https://www.amazon.com/ | https://www.amazon.com/ap/signin?_encoding=UTF8&openid.assoc_handle=amazon_checkout_us&openid.claimed_id=http%3A%2F%2Fspecs.openid.net%2Fauth%2F2.0%2Fidentifier_select&openid.identity=http%3A%2F%2Fspecs.openid.net%2Fauth%2F2.0%2Fidentifier_select&openid.mode=checkid_setup&openid.ns=http%3A%2F%2Fspecs.openid.net%2Fauth%... | 13 | keep | true | true | Agent searched for winter sale stuffed animals, sorted by price low-to-high, added item to cart, and proceeded to checkout as instructed. | claude-haiku-4-5-20251001 | page: https://www.amazon.com/
1. click [184] where [184] is Search Amazon
2. type [184] [Winter Sale toys ] where [184] is Search Amazon
page: https://www.amazon.com/s?k=Winter+Sale+toys&crid=15FG1S7D9BO87&sprefix=winter+sale+toys%2Caps%2C131&ref=nb_sb_noss
3. hover [86] where [86] is
4. type [86] [Toy Deals] whe... |
openweb_1039 | train | Find high-severity advisories on GitHub. | https://github.com/ | https://github.com/advisories?page=2&query=severity%3Ahigh | 6 | keep | true | true | Objective is clear and well-formed. Trajectory coherently navigates to GitHub Advisory Database, filters by high severity, and pages to results matching the objective. | claude-haiku-4-5-20251001 | page: https://github.com/
1. goto [https://github.com/security]
page: https://github.com/security
2. click [703] where [703] is GitHub Advisory Database
page: https://github.com/advisories
3. click [653] where [653] is Severity
4. click [672] where [672] is High
page: https://github.com/advisories?query=sev... |
openweb_104 | train | Get information on machine learning and related topics, possibly focusing on natural language processing. | https://arxiv.org/ | http://arxiv.org/search/?query=machine+learning&searchtype=all&source=header | 9 | keep | true | true | Clear objective to explore machine learning and NLP on arXiv. Trajectory searches for both topics and examines relevant papers appropriately. | claude-haiku-4-5-20251001 | page: https://arxiv.org/
1. type [50] [machine learning ] where [50] is Search term or terms
page: http://arxiv.org/search/?query=machine+learning&searchtype=all&source=header
2. click [205] where [205] is arXiv:2412.18601
page: https://arxiv.org/abs/2412.18601
3. click [219] where [219] is
4. click [276] wher... |
openweb_1040 | train | Find high-severity GHSA advisories related to GitHub Actions. | https://github.com/ | https://github.com/advisories?query=severity%3AHigh+type%3Areviewed+ecosystem%3Aactions | 9 | drop | true | false | Trajectory ends at correct URL but agent gives up prematurely, concluding no results exist when none actually found. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [1178] where [1178] is Security
page: https://github.com/security
2. click [703] where [703] is GitHub Advisory Database
page: https://github.com/advisories
3. click [641] where [641] is Search by CVE/GHSA ID, package, severity, ecosystem, credit…
4. type [641] [severity:High... |
openweb_1041 | train | Browse the unreviewed high severity security advisories on GitHub. | https://github.com/ | https://github.com/advisories?query=severity%3AHigh+type%3Aunreviewed | 13 | keep | true | true | Agent navigates to GitHub advisories page, searches for high severity unreviewed advisories, reaching the correct endpoint URL with appropriate filters applied. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [1178] where [1178] is Security
page: https://github.com/security
2. click [703] where [703] is GitHub Advisory Database
page: https://github.com/advisories
3. click [641] where [641] is Search by CVE/GHSA ID, package, severity, ecosystem, credit…
4. type [641] [severity:High... |
openweb_1042 | train | Investigate a specific security vulnerability (e.g., CVE-2022-48470) and gather detailed information about its impact and CWE. | https://github.com/ | https://nvd.nist.gov/vuln/detail/CVE-2022-48470 | 17 | keep | true | true | Agent navigates GitHub Security and NVD to research CVE-2022-48470, successfully reaching the target NVD page and exploring CWE details. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [1178] where [1178] is Security
page: https://github.com/security
2. click [703] where [703] is GitHub Advisory Database
page: https://github.com/advisories
3. click [641] where [641] is Search by CVE/GHSA ID, package, severity, ecosystem, credit…
4. type [641] [severity:High... |
openweb_1043 | train | Find information about a specific vulnerability with high severity on GitHub. | https://github.com/ | https://github.com/advisories/GHSA-vp3x-9qpg-q385 | 7 | keep | true | true | Objective is clear: find high-severity vulnerability info on GitHub. Trajectory successfully navigates to GitHub Advisory Database, filters by high severity, and lands on a specific vulnerability advisory page. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [289] where [289] is Resources
2. click [301] where [301] is Security
page: https://github.com/resources/articles/security
3. click [728] where [728] is Security
page: https://github.com/security
4. click [703] where [703] is GitHub Advisory Database
page: https://github.c... |
openweb_1044 | train | Investigate high-severity security advisories on GitHub. | https://github.com/ | https://github.com/advisories?page=2&query=severity%3Ahigh | 7 | keep | true | true | Objective is clear and coherent. Trajectory systematically navigates to GitHub's Advisory Database, filters for high-severity advisories, and advances to the next page as instructed. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [163] where [163] is Navigation Menu
2. click [1177] where [1177] is Security
page: https://github.com/security
3. click [701] where [701] is GitHub Advisory Database
page: https://github.com/advisories
4. click [653] where [653] is Severity
5. click [672] where [672] is High... |
openweb_1045 | train | Find all high-severity security advisories on GitHub. | https://github.com/ | https://github.com/advisories?page=2&query=severity%3Ahigh | 6 | keep | true | true | Trajectory directly pursues finding high-severity security advisories by navigating to GitHub Advisory Database, filtering by high severity, and viewing results across pages. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [1178] where [1178] is Security
page: https://github.com/security
2. click [701] where [701] is GitHub Advisory Database
page: https://github.com/advisories
3. click [653] where [653] is Severity
4. click [672] where [672] is High
page: https://github.com/advisories?query=... |
openweb_1046 | train | Find high-severity security vulnerabilities related to the RabbitMQ project and report a vulnerability. | https://github.com/ | https://github.com/login?return_to=https%3A%2F%2Fgithub.com%2Frabbitmq%2Frabbitmq-server%2Fsecurity%2Fadvisories%2Fnew | 33 | keep | true | true | Objective is clear: find RabbitMQ high-severity vulnerabilities and report one. Trajectory explores GitHub advisories, locates RabbitMQ vulnerabilities, and navigates to the vulnerability reporting form. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [1178] where [1178] is Security
page: https://github.com/security
2. click [703] where [703] is GitHub Advisory Database
page: https://github.com/advisories
3. click [641] where [641] is Search by CVE/GHSA ID, package, severity, ecosystem, credit…
4. type [641] [severity:High... |
openweb_1047 | train | Find the scores of college football games. | https://www.espn.com/ | https://www.espn.com/college-football/game/_/gameId/401677101/iowa-missouri | 5 | keep | true | true | Agent searches for college football scores and navigates to game pages displaying scores. The trajectory directly pursues this objective despite not finding final scores. | claude-haiku-4-5-20251001 | page: https://www.espn.com/
1. type [1109] [NBA scores ] where [1109] is Search
2. click [218] where [218] is OU 20 NAVY 21
page: https://www.espn.com/college-football/boxscore/_/gameId/401677090
3. scroll [down]
4. click [3662] where [3662] is IOWA 8-4 19 MIZ 9-3 Mon- 11:30 AM ESPN
page: https://www.espn.com/... |
openweb_1049 | train | Find college football scores. | https://www.espn.com/ | https://www.espn.com/college-football/scoreboard | 3 | keep | true | true | Objective is clear and well-formed. Trajectory directly pursues finding college football scores, reaching the scoreboard page via ESPN's CFB section. | claude-haiku-4-5-20251001 | page: https://www.espn.com/
1. click [4780] where [4780] is CFB Bowl Games
page: https://www.espn.com/college-football/story/_/page/cfpbowls2024/college-football-bowl-2024-2025-schedule-matchups
2. click [2553] where [2553] is Scores
page: https://www.espn.com/college-football/scoreboard
3. stop ["Found"] (fi... |
openweb_1050 | train | Search for pre-trained models on Hugging Face | https://huggingface.co/ | https://huggingface.co/sophia-jihye/Incorporation_of_Company-Related_Factual_Knowledge_into_Pre-trained_Language_Models | 5 | keep | true | true | Agent searches for pre-trained models on Hugging Face via Models section and search box, successfully finding a specific pre-trained model despite not displaying full search results list. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. click [52] where [52] is Models
page: https://huggingface.co/models
2. click [78] where [78] is Search models, datasets, users...
3. type [78] [pre-trained models ] where [78] is Search models, datasets, users...
page: https://huggingface.co/sophia-jihye/Incorporation_of_Compan... |
openweb_1051 | train | Look up models that incorporate company-related factual knowledge. | https://huggingface.co/ | https://huggingface.co/sophia-jihye/Incorporation_of_Company-Related_Factual_Knowledge_into_Pre-trained_Language_Models | 9 | keep | true | true | Objective to find models incorporating company knowledge is coherent. Trajectory searches HuggingFace models and lands on SEC-BERT, a model specifically designed for company-related factual knowledge integration. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. click [52] where [52] is Models
page: https://huggingface.co/models
2. click [78] where [78] is Search models, datasets, users...
3. type [78] [pre-trained models ] where [78] is Search models, datasets, users...
page: https://huggingface.co/sophia-jihye/Incorporation_of_Compan... |
openweb_1052 | train | Find stays in San Francisco. | https://www.booking.com/ | https://www.booking.com/city/us/san-francisco.html | 4 | keep | true | true | Clear, coherent objective. Trajectory successfully navigates to San Francisco stays page on Booking.com despite final parsing error. | claude-haiku-4-5-20251001 | page: https://www.booking.com/
1. click [4186] where [4186] is Dismiss sign-in info.
2. click [329] where [329] is Where are you going?
3. type [329] [San Francisco ] where [329] is Where are you going?
page: https://www.booking.com/city/us/san-francisco.html
4. assistant Cannot correctly parse the action (final... |
openweb_1060 | train | Find the model "Music-AI-Voices" for game characters AI on Hugging Face. | https://huggingface.co/ | https://huggingface.co/QuickWick/Music-AI-Voices | 2 | keep | true | true | Objective is clear and coherent. Trajectory searches for the model and reaches the correct page, matching the stated goal. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. type [44] [Music-AI-Voices ] where [44] is Search models, datasets, users...
page: https://huggingface.co/QuickWick/Music-AI-Voices
2. stop [Music-AI-Voices for game characters] (final recorded decision) |
openweb_1061 | train | How are downloads counted for models | https://huggingface.co/ | https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/model-libraries.ts | 13 | keep | true | true | Agent navigates HF models site, finds download stats documentation, and reaches GitHub code implementing download counting as requested. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. click [51] where [51] is
page: https://huggingface.co/models
2. type [78] [game character AI ] where [78] is Search models, datasets, users...
page: https://huggingface.co/search/full-text?q=game%20character%20AI
3. scroll [down]
4. click [98] where [98] is
page: https://hug... |
openweb_1063 | train | Find out the pricing and features of GitHub Copilot. | https://huggingface.co/ | https://www.google.com/search?q=GitHub+Copilot+pricing+and+features&sca_esv=768f8b678b2bd642&source=hp&ei=CYN7Z_O4EcjDkPIPhqKD8Qs&iflsig=AL9hbdgAAAAAZ3uRGbHDuAleB-LWqGIrUi7o6lvMQiKw&ved=0ahUKEwiz4dGhxuCKAxXIIUQIHQbRIL4Q4dUDCBA&uact=5&oq=GitHub+Copilot+pricing+and+features&gs_lp=Egdnd3Mtd2l6IiNHaXRIdWIgQ29waWxvdCBwcmlja... | 4 | keep | true | true | Objective is clear and well-formed. Trajectory searches for the requested information and identifies a pricing detail ($10/month), directly pursuing the stated goal. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. new_tab
2. goto [https://www.google.com]
page: https://www.google.com/
3. type [95] [GitHub Copilot pricing and features ] where [95] is Search
page: https://www.google.com/search?q=GitHub+Copilot+pricing+and+features&sca_esv=768f8b678b2bd642&source=hp&ei=CYN7Z_O4EcjDk…
4. stop... |
openweb_1064 | train | Find out the cost of GitHub Copilot for individual developers. | https://huggingface.co/ | https://github.com/pricing | 4 | keep | true | true | Objective is clear and coherent. Trajectory navigates to GitHub pricing page and finds Copilot pricing information ($10/month or $100/year) as instructed. | claude-haiku-4-5-20251001 | page: https://huggingface.co/
1. goto [github.com]
2. goto [https://github.com/pricing]
page: https://github.com/pricing
3. click [3051] where [3051] is What is the pricing for GitHub Copilot?
4. stop [$10 USD/month or $100 USD/year] (final recorded decision) |
openweb_1067 | heldout | Find the Instant Pot Crispy Chicken Carnitas meal prep recipe. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/274921/instant-pot-crispy-chicken-carnitas/ | 3 | keep | true | true | Objective is clear and well-formed. Trajectory successfully searches for and navigates to the requested recipe page. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [881] [Instant Pot Crispy Chicken Carnitas meal prep ] where [881] is Search the site
page: https://www.allrecipes.com/search?q=Instant+Pot+Crispy+Chicken+Carnitas+meal+prep
2. click [666] where [666] is Save Recipe MEXICAN Instant Pot® Crispy Chicken Carnitas 14 Ratings
... |
openweb_1069 | heldout | Find and print a meal prep recipe. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/264155/mango-chicken-meal-prep-bowls/ | 4 | keep | true | true | Objective is clear. Trajectory searches for meal prep recipes on AllRecipes, finds 'Mango Chicken Meal Prep Bowls,' and prints it as requested. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [882] [meal prep ] where [882] is Search the site
page: https://www.allrecipes.com/search?q=meal+prep
2. click [695] where [695] is Save Recipe CHICKEN BREAST Mango Chicken Meal Prep Bowls 35 Ratings
page: https://www.allrecipes.com/recipe/264155/mango-chicken-meal-pre... |
openweb_1070 | heldout | Find a healthier version of chicken enchiladas meal prep recipe and print it. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/274921/instant-pot-crispy-chicken-carnitas/?print | 17 | drop | true | false | Trajectory ends on chicken carnitas recipe, not enchiladas. Agent found healthier enchiladas but printed carnitas instead. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [892] [meal prep recipes ] where [892] is Search the site
page: https://www.allrecipes.com/search?q=meal+prep+recipes
2. click [667] where [667] is Asian Chicken Salad in a Jar MEAL PREP RECIPES 20 Meal Prep Chicken Recipes That Aren't Boring
page: https://www.allrecip... |
openweb_1072 | train | Find a Python data science course taught by a reputable institution. | https://www.coursera.org/ | https://www.coursera.org/search?query=Python%20data%20science | 2 | keep | true | true | Objective is clear. Trajectory searches for Python data science courses on Coursera, a reputable institution platform, reaching search results. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. type [2058] [Python data science ] where [2058] is What do you want to learn?
page: https://www.coursera.org/search?query=Python%20data%20science
2. stop [https://www.coursera.org/learn/python-data] (final recorded decision) |
openweb_1073 | train | Find hotels in SoHo, New York City. | https://www.google.com/maps | https://www.google.com/maps/place/SoHo,+New+York,+NY/@40.6970193,-74.3093286,10z/data=!4m6!3m5!1s0x89c2598c5e51e2f3:0xce2f3449f490f818!8m2!3d40.7245908!4d-74.0019174!16zL20vMHBuMjA?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 5 | keep | true | true | Agent successfully navigated to SoHo neighborhood on Google Maps and identified hotels there, directly pursuing the stated objective. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [139] [New York City ] where [139] is Search Google Maps
page: https://www.google.com/maps/place/New+York,+NY/@40.6970193,-74.3093286,10z/data=!3m1!4b1!4m6!3m5!1s0x89c24fa5d33f083b:0…
2. click [677] where [677] is Hotels
3. click [749] where [749] is Neighborhoods
4. clic... |
openweb_1074 | train | Find hotels in the SoHo neighborhood of New York City. | https://www.google.com/maps | https://www.google.com/maps/search/hotels+in+SoHo+New+York+City/@40.7229874,-74.0105525,15z/data=!3m1!4b1?entry=ttu&g_ep=EgoyMDI1MDExMC4wIKXMDSoASAFQAw%3D%3D | 3 | keep | true | true | Objective is clear and coherent. Trajectory directly searches for hotels in SoHo NYC on Google Maps, achieving the goal successfully. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. click [142] where [142] is Search Google Maps
page: https://www.google.com/maps/@37.4333076,-122.1820416,14z?entry=ttu&g_ep=EgoyMDI1MDExMC4wIKXMDSoASAFQAw%3D%3D
2. type [142] [hotels in SoHo New York City ] where [142] is Search Google Maps
page: https://www.google.com/maps... |
openweb_1075 | train | Find affordable accommodations in SoHo, New York City for 2 guests. | https://www.google.com/maps | https://www.google.com/maps/search/Hotels/@40.7236406,-74.0191332,15z/data=!3m1!4b1!4m7!2m6!5m4!5m3!1s2025-01-08!4m1!1i2!6e3?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 13 | keep | true | true | Agent searches SoHo hotels for 2 guests, filters dates and price. Trajectory pursues the stated objective despite not completing the full search. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [139] [New York City ] where [139] is Search Google Maps
page: https://www.google.com/maps/place/New+York,+NY/@40.6970193,-74.3093286,10z/data=!3m1!4b1!4m6!3m5!1s0x89c24fa5d33f083b:0…
2. click [677] where [677] is Hotels
3. click [749] where [749] is Neighborhoods
4. clic... |
openweb_1077 | train | Book a hotel in New York City from January 7 to January 9 for 2 guests within a budget of $160 with free cancellation. | https://www.google.com/maps | https://www.google.com/maps/place/Hyatt+House+Jersey+City/@40.7158618,-74.0519954,15z/data=!4m24!1m10!2m9!1sHotels!5m6!5m5!1s2025-01-07!2i2!4m1!1i2!5b1!6e3!3m12!1s0x89c251ac2f168bc5:0x5056d338fd868102!5m5!1s2025-01-07!2i2!4m1!1i2!5b1!8m2!3d40.7158618!4d-74.033971!15sCgZIb3RlbHOSAQVob3RlbOABAA!16s%2Fg%2F11r9lczjdy?entry... | 21 | drop | true | false | Trajectory never actually books a hotel; it only searches and views details. Booking is the core objective requirement. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [139] [New York City ] where [139] is Search Google Maps
page: https://www.google.com/maps/place/New+York,+NY/@40.6970193,-74.3093286,10z/data=!3m1!4b1!4m6!3m5!1s0x89c24fa5d33f083b:0…
2. click [677] where [677] is Hotels
3. click [749] where [749] is Neighborhoods
4. clic... |
openweb_1078 | train | Find a place to stay in New York City for a short trip. | https://www.google.com/maps | https://www.booking.com/hotel/us/1850s-historic-row-house-7min-train-to-nyc-with-private-backyard.html | 25 | keep | true | true | Objective clearly requests finding NYC accommodation for short trip. Trajectory systematically searches Google Maps hotels, filters options, and ends on Booking.com property listing—directly pursuing the goal. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [139] [New York City ] where [139] is Search Google Maps
page: https://www.google.com/maps/place/New+York,+NY/@40.6970193,-74.3093286,10z/data=!3m1!4b1!4m6!3m5!1s0x89c24fa5d33f083b:0…
2. click [677] where [677] is Hotels
3. click [749] where [749] is Neighborhoods
4. clic... |
openweb_1079 | train | Find a place to stay in New York City. | https://www.google.com/maps | https://www.google.com/maps/place/Hampton+Inn+Manhattan%2FTimes+Square+Central/@40.7400387,-74.0319234,13z/data=!4m13!1m2!2m1!1shotels+near+New+York+City!3m9!1s0x89c259ab5432861d:0x6f25cb902cee5cb1!5m2!4m1!1i2!8m2!3d40.7556071!4d-73.9886633!15sChlob3RlbHMgbmVhciBOZXcgWW9yayBDaXR5kgEFaG90ZWzgAQA!16s%2Fg%2F11bcf3p3b3?ent... | 5 | keep | true | true | Objective is clear: find NYC accommodation. Trajectory searches hotels on Google Maps, selects Hampton Inn, and checks availability and rooms. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [hotels near New York City ] where [142] is Search Google Maps
page: https://www.google.com/maps/search/hotels+near+New+York+City/@40.7400387,-74.0319234,13z/data=!3m1!4b1?entry=ttu&g_ep=E…
2. click [840] where [840] is Hampton Inn Manhattan/Times Square Central
... |
openweb_1081 | heldout | Find fruit dessert recipes. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipes/17140/desserts/fruit-desserts/ | 5 | keep | true | true | The objective is clear and achievable. The trajectory successfully navigates to a fruit desserts recipe page using appropriate search and navigation steps. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. click [889] where [889] is Search the site
2. type [889] [desserts ] where [889] is Search the site
page: https://www.allrecipes.com/search?q=desserts
3. type [122] [fruit desserts ] where [122] is Search the site
page: https://www.allrecipes.com/search?q=fruit+desserts
4. ... |
openweb_1082 | heldout | Find an apple dessert recipe. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/18359/apple-enchilada-dessert/ | 3 | keep | true | true | User searches for apple dessert recipe and reaches an apple dessert recipe page. Clear objective pursuit with appropriate site navigation. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [885] [apple dessert ] where [885] is Search the site
page: https://www.allrecipes.com/search?q=apple+dessert
2. goto [https://www.allrecipes.com/recipe/18359/apple-enchilada-dessert/]
page: https://www.allrecipes.com/recipe/18359/apple-enchilada-dessert/
3. stop [Appl... |
openweb_1083 | heldout | Find apple dessert recipes. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/59124/apple-turnovers/ | 13 | keep | true | true | The objective is clear: find apple dessert recipes. The trajectory systematically searches for and navigates through multiple apple dessert recipes on AllRecipes, ending at apple turnovers. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. click [889] where [889] is Search the site
2. type [889] [desserts ] where [889] is Search the site
page: https://www.allrecipes.com/search?q=desserts
3. type [122] [fruit desserts ] where [122] is Search the site
page: https://www.allrecipes.com/search?q=fruit+desserts
4. ... |
openweb_1084 | train | Get information about the latest iPhone models. | https://www.apple.com/ | https://www.apple.com/shop/buy-iphone | 4 | keep | true | true | Objective is clear. Trajectory navigates Apple site and reaches iPhone purchase page displaying latest models (iPhone 16 series), plausibly pursuing the information goal. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [170] where [170] is Global
page: https://www.apple.com/airpods/
2. click [171] where [171] is
page: https://www.apple.com/store
3. click [915] where [915] is
page: https://www.apple.com/shop/buy-iphone
4. stop [iPhone 16 Pro, iPhone 16 Pro Max, iPhone 16, and iPhone 16... |
openweb_1085 | train | Find the price of the 256 GB iPad Air for education. | https://www.apple.com/ | https://www.apple.com/us-edu/shop/buy-ipad/ipad-air | 9 | keep | true | true | Trajectory successfully navigates to iPad Air education pricing page and clicks 256GB option. Goal clearly pursued and found. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [1378] where [1378] is iPhone
page: https://www.apple.com/iphone/
2. click [1013] where [1013] is Shop iPhone
page: https://www.apple.com/shop/buy-iphone
3. click [977] where [977] is
page: https://www.apple.com/shop/buy-iphone/iphone-16-pro
4. click [118] where [118] i... |
openweb_1087 | train | Find trips to Japan from January 1 to January 3, 2025. | https://www.google.com/travel/flights | https://www.google.com/travel/flights?tfs=CBwQARocagwIAhIIL20vMGQ2bHByDAgEEggvbS8wM18zZBocagwIBBIIL20vMDNfM2RyDAgCEggvbS8wZDZscEABSAFwAYIBCwj___________8BmAEB&tfu=KgIIAw | 9 | keep | true | true | Objective is clear and coherent. Trajectory coherently pursues finding Japan trips for the specified dates on Google Flights, selecting destination and dates despite incomplete search results. | claude-haiku-4-5-20251001 | page: https://www.google.com/travel/flights
1. click [638] where [638] is Where to?
2. click [1626] where [1626] is Japan
page: https://www.google.com/travel/flights?tfs=CBwQARocagwIAhIIL20vMGQ2bHByDAgEEggvbS8wM18zZBocagwIBBIIL20vMDNfM2RyDAgCEggvb…
3. click [691] where [691] is Departure
4. click [1810] where [18... |
openweb_1089 | train | Find out how to build a custom integration with GitHub using their GraphQL API. | https://github.com/ | https://docs.github.com/en/graphql | 9 | keep | true | true | Objective is clear and coherent. Trajectory navigates from GitHub homepage to GraphQL API documentation, directly pursuing the stated goal of learning to build custom GitHub integrations using GraphQL. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [158] where [158] is Close
2. click [601] where [601] is Pause demo
3. click [183] where [183] is
4. click [1171] where [1171] is
5. Attempt to perfom "click" on element "[280]" but no matching element found. Please check the observation more carefully.
6. click [1172] where [1172]... |
openweb_1092 | train | What are the current NCAAF scores. | https://www.espn.com/ | https://www.espn.com/college-football/scoreboard | 4 | keep | true | true | Clear objective to find NCAAF scores. Trajectory navigates ESPN and reaches scoreboard page showing current scores. | claude-haiku-4-5-20251001 | page: https://www.espn.com/
1. scroll [down]
2. click [3862] where [3862] is
page: https://www.espn.com/college-football/story/_/page/cfpbowls2024/college-football-bowl-2024-2025-schedule-matchups
3. click [2548] where [2548] is Scores
page: https://www.espn.com/college-football/scoreboard
4. stop [Available o... |
openweb_1093 | train | Get the stats and recent updates for the Dallas Mavericks after their game against the Phoenix Suns. | https://www.google.com/ | https://www.espn.com/nba/team/_/name/dal/dallas-mavericks | 5 | keep | true | true | The trajectory coherently pursues the objective: searches NBA scores, finds the Mavericks-Suns game result, and navigates to the Mavericks team stats page. | claude-haiku-4-5-20251001 | page: https://www.google.com/
1. type [109] [NBA scores ] where [109] is Search
page: https://www.google.com/search?q=NBA+scores&sca_esv=93ff44d7863fe196&source=hp&ei=i99vZ8fjBf3ekPIP2fWByQo&iflsig=AL9hbdg…
2. click [591] where [591] is NEWS
page: https://www.google.com/search?q=NBA+scores&sca_esv=93ff44d7863f... |
openweb_1094 | train | Compare the performance of the Dallas Mavericks and the Cleveland Cavaliers for the 2024-2025 NBA season. | https://www.google.com/ | https://www.google.com/search?q=%22Dallas+Mavericks+vs+Cleveland+Cavaliers+2024-2025+NBA+season%22&sca_esv=768f8b678b2bd642&source=hp&ei=toV7Z-WfD8Gwur8P-qbI2Qs&iflsig=AL9hbdgAAAAAZ3uTxvo6nYYUyAG97BAb9Fj7yeJiUgzW&ved=0ahUKEwjl0qDoyOCKAxVBmO4BHXoTMrsQ4dUDCA4&uact=5&oq=%22Dallas+Mavericks+vs+Cleveland+Cavaliers+2024-2025... | 2 | keep | true | true | User asks to compare two NBA teams' 2024-2025 performance. Trajectory searches for this comparison and stops after finding relevant results. | claude-haiku-4-5-20251001 | page: https://www.google.com/
1. type [93] ["Dallas Mavericks vs Cleveland Cavaliers 2024-2025 NBA season" ] where [93] is Search
page: https://www.google.com/search?q=%22Dallas+Mavericks+vs+Cleveland+Cavaliers+2024-2025+NBA+season%22&sca_esv=768f8b678b2b…
2. stop ["Cleveland Cavaliers winermalinkedly with a reco... |
openweb_1098 | train | Find out the price of Apple AirPods. | https://www.apple.com/ | https://www.apple.com/airpods/ | 2 | keep | true | true | Clear objective to find AirPods price. Trajectory directly navigates to AirPods page and stops at price display ($129). | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. goto [https://www.apple.com/airpods/]
page: https://www.apple.com/airpods/
2. stop [$129] (final recorded decision) |
openweb_1099 | train | Compare AirPods models available for purchase. | https://www.apple.com/ | https://www.apple.com/airpods/compare/ | 9 | keep | true | true | The objective is clear and coherent. The trajectory successfully navigates to the AirPods comparison page, visiting relevant product pages and ending at the correct URL displaying AirPods model comparisons. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [954] where [954] is Learn more, iPhone 16 Pro
page: https://www.apple.com/iphone-16-pro/
2. click [1089] where [1089] is A hand holds iPhone 16 Pro in Desert Titanium, all screen display, Dynamic Island centered near top, rounded corners
3. click [1865] where [1865] is Learn... |
openweb_11 | train | Find a BBC News video. | https://www.bbc.com/news | https://www.bbc.com/news/videos/c70e2g09ng9o | 4 | keep | true | true | Objective is clear and well-formed. Trajectory successfully navigates to BBC News videos and finds a specific video, reaching the end URL. | claude-haiku-4-5-20251001 | page: https://www.bbc.com/news
1. goto [https://www.bbc.com/news/live/c5y817zzgxqt]
page: https://www.bbc.com/news/live/c5y817zzgxqt
2. click [266] where [266] is Video
page: https://www.bbc.com/news/videos/c4gdgl3e404o
3. click [833] where [833] is Split screen of a man and a driverless car Man trapped inside... |
openweb_110 | train | Find hotels near Mt Everest available on January 14. | https://www.google.com/maps | https://www.google.com/maps/search/Hotels/@27.9880925,86.8425734,12z/data=!3m1!4b1!4m15!2m14!3m6!1sHotels!2sMt+Everest!3s0x39e854a215bd9ebd:0x576dcf806abbab2!4m2!1d86.9249751!2d27.9881206!5m5!5m4!1s2025-01-15!2i2!4m1!1i2!6e3?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 5 | drop | true | false | Objective asks for January 14, but trajectory sets check-in to January 15. The recorded actions don't match the specified date. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [Mount Everest ] where [142] is Search Google Maps
page: https://www.google.com/maps/place/Mt+Everest/@27.9881186,86.9043755,14z/data=!3m1!4b1!4m6!3m5!1s0x39e854a215bd9ebd:0x57…
2. click [693] where [693] is Search nearby Mt Everest
page: https://www.google.com/m... |
openweb_1100 | train | Find open-source repositories that use JavaScript. | https://github.com/ | https://github.com/search?q=language%3Ajavascript&type=repositories&p=6 | 7 | keep | true | true | Agent searches GitHub for JavaScript repositories using language filter, navigating multiple result pages. Objective and trajectory align well. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [383] where [383] is Search or jump to…
2. type [406] [language:javascript ] where [406] is Search
page: https://github.com/search?q=language%3Ajavascript&type=repositories
3. click [1283] where [1283] is Next Page
page: https://github.com/search?q=language%3Ajavascript&type=... |
openweb_1101 | train | Find a popular JavaScript project on GitHub. | https://github.com/ | https://github.com/search?q=javascript+language%3Ajavascript&type=repositories&p=4 | 6 | keep | true | true | Agent searches for JavaScript projects on GitHub and navigates to results pages, stopping at a popular project (Node.js with 108k stars). Trajectory aligns with objective. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [382] where [382] is Search or jump to…
2. type [405] [javascript language:javascript ] where [405] is Search
page: https://github.com/search?q=javascript%20language%3Ajavascript&type=repositories
3. click [1355] where [1355] is Next Page
page: https://github.com/search?q=jav... |
openweb_1102 | train | Find information about the Kotlin programming language on GitHub. | https://github.com/ | https://github.com/search?q=Kotlin&type=repositories | 14 | keep | true | true | Objective is clear: find Kotlin info on GitHub. Trajectory searches for Kotlin on GitHub despite detours to status page and Copilot page, ultimately reaching search results. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [383] where [383] is Search or jump to…
2. type [406] [Kotlin ] where [406] is Search
page: https://github.com/search?q=Kotlin&type=repositories
3. click [17] where [17] is GitHub Status
page: https://www.githubstatus.com/
4. click [38] where [38] is GitHub Octicon logo
5. cl... |
openweb_1103 | train | Investigate a specific workflow (Close stale PRs) of the kotlin-web-site GitHub repository and retrieve its yaml file. | https://github.com/ | https://github.com/JetBrains/kotlin-web-site/blob/master/.github/workflows/close-stale-pr.yml | 5 | keep | true | true | Objective is clear and well-formed. Trajectory navigates to the kotlin-web-site repository, accesses GitHub Actions, finds the specific workflow, and retrieves its YAML file content as instructed. | claude-haiku-4-5-20251001 | page: https://github.com/
1. goto [https://github.com/JetBrains/kotlin-web-site]
page: https://github.com/JetBrains/kotlin-web-site
2. click [611] where [611] is Actions
page: https://github.com/JetBrains/kotlin-web-site/actions
3. goto [https://github.com/JetBrains/kotlin-web-site/actions/workflows/close-stal... |
openweb_1105 | train | View the course offerings in different categories on Coursera. | https://www.coursera.org/ | https://www.coursera.org/browse/language-learning | 17 | keep | true | true | Objective is clear: view course offerings across categories. Trajectory systematically explores multiple categories (Business, Computer Science, Health, Arts/Humanities, Language Learning), demonstrating the requested behavior despite some navigation detours. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. scroll [down]
2. scroll [down]
3. click [3213] where [3213] is Business 1095 courses
4. click [3220] where [3220] is Computer Science 668 courses
page: https://www.coursera.org/browse/business
5. type [2072] [Business name] [press_enter_after=1 ] where [2072] is What do you want... |
openweb_1106 | train | Explore Coursera's Arts and Humanities courses, and specifically history courses. | https://www.coursera.org/ | https://www.coursera.org/browse/arts-and-humanities/history | 9 | keep | true | true | Agent explores Coursera's Arts/Humanities courses and reaches history courses. Initial detours through Data Science and Health are normal exploration. Final destination matches objective. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. click [3201] where [3201] is Data Science 425 courses
2. click [3222] where [3222] is Health 471 courses
page: https://www.coursera.org/browse/data-science
3. click [4594] where [4594] is Other topics to explore
4. click [4597] where [4597] is Arts and Humanities 338 courses
... |
openweb_1107 | train | Search for a course on ancient Greek history on Coursera. | https://www.coursera.org/ | https://www.coursera.org/learn/ancient-greeks | 13 | keep | true | true | Clear objective to search for ancient Greek history course on Coursera. Trajectory navigates through categories, searches for relevant terms, and successfully reaches the target course page. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. click [3201] where [3201] is Data Science 425 courses
2. click [3222] where [3222] is Health 471 courses
page: https://www.coursera.org/browse/data-science
3. click [4594] where [4594] is Other topics to explore
4. click [4597] where [4597] is Arts and Humanities 338 courses
... |
openweb_111 | train | Find reviews for hotels near Mt Everest. | https://www.google.com/maps | https://www.google.com/maps/place/Mountain+Lodges+of+Nepal+-+Namche/@27.804674,86.5685992,12z/data=!4m31!1m16!2m15!1sHotels!3m6!1sHotels!2sMt+Everest!3s0x39e854a215bd9ebd:0x576dcf806abbab2!4m2!1d86.9249751!2d27.9881206!5m5!5m4!1s2025-01-15!2i3!4m1!1i2!6e3!3m13!1s0x39e908c5681ba72d:0x68b397779d598bd0!5m4!1s2025-01-15!2i... | 9 | keep | true | true | Agent successfully searched for hotels near Mt Everest on Google Maps and accessed reviews for a specific hotel (Mountain Lodges of Nepal - Namche), directly fulfilling the objective. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [Mount Everest ] where [142] is Search Google Maps
page: https://www.google.com/maps/place/Mt+Everest/@27.9881186,86.9043755,14z/data=!3m1!4b1!4m6!3m5!1s0x39e854a215bd9ebd:0x57…
2. click [693] where [693] is Search nearby Mt Everest
page: https://www.google.com/m... |
openweb_1111 | heldout | Define box office. | https://dictionary.cambridge.org/ | https://dictionary.cambridge.org/dictionary/english/box-office | 2 | keep | true | true | Clear objective to define a term. Trajectory searches for 'box office' on Cambridge Dictionary and reaches the definition page, successfully completing the task. | claude-haiku-4-5-20251001 | page: https://dictionary.cambridge.org/
1. type [272] [box office ] where [272] is Search
page: https://dictionary.cambridge.org/dictionary/english/box-office
2. stop [the place in a cinema or theatre where tickets are sold; the number of tickets sold for a film, as a measure of how successful a film or actor is]... |
openweb_1112 | train | Find out the price of GitHub Copilot. | https://github.com/ | https://github.com/features/copilot | 2 | keep | true | true | Agent navigated directly to GitHub Copilot pricing page and extracted clear pricing information matching the objective. | claude-haiku-4-5-20251001 | page: https://github.com/
1. goto [https://github.com/features/copilot]
page: https://github.com/features/copilot
2. stop [$0/ Free, $10/ Pro, $19/ Business, $39/ Enterprise] (final recorded decision) |
openweb_1113 | train | Find out the price of GitHub Copilot. | https://github.com/ | https://github.com/login?return_to=%2Fgithub-copilot%2Fsignup%3Fcft%3Dcopilot_lo.features_copilot.cfi&cft=copilot_lo.features_copilot.cfi | 9 | drop | true | false | Trajectory stops at login page before revealing pricing information. Agent never accesses actual pricing details. | claude-haiku-4-5-20251001 | page: https://github.com/
1. click [158] where [158] is Close
2. click [184] where [184] is
3. click [191] where [191] is
page: https://github.com/features/copilot
4. Attempt to perfom "type" on element "[191]" but no matching element found. Please check the observation more carefully.
5. click [959] where [959] ... |
openweb_1114 | train | The most likely instruction is to "Get the features of the iPhone 16 Pro Max." | https://www.apple.com/ | https://www.apple.com/iphone/ | 11 | drop | true | false | Trajectory visits iPhone 16 Pro page but doesn't systematically explore features. Multiple detours and backtracking suggest unclear navigation toward the goal. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [190] where [190] is Store
page: https://www.apple.com/store
2. goto [https://www.apple.com/shop/buy-iphone/iphone-16-pro]
page: https://www.apple.com/shop/buy-iphone/iphone-16-pro
3. go_back
page: https://www.apple.com/store
4. scroll [down]
5. click [900] where [900] ... |
openweb_1115 | train | Find out about the camera features of the iPhone 16 Pro. | https://www.apple.com/ | https://www.apple.com/iphone-16-pro/ | 2 | keep | true | true | Objective is clear and coherent. Trajectory navigates to iPhone 16 Pro page and stops after finding camera information, directly addressing the objective. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [954] where [954] is Learn more, iPhone 16 Pro
page: https://www.apple.com/iphone-16-pro/
2. stop [Camera features: Take total Camera Control, 4K 120 fps Dolby Vision, New 48MP Ultra Wide camera, and more.] (final recorded decision) |
openweb_1116 | train | Does the iPhone have a 28 mm camera? | https://www.apple.com/ | https://www.apple.com/iphone-16-pro/ | 13 | drop | false | false | Objective is a yes/no question, not an instruction. Trajectory explores iPhone cameras but never answers the specific 28mm question. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [188] where [188] is
page: https://www.apple.com/store
2. click [305] where [305] is
page: https://www.apple.com/iphone/
3. click [2359] where [2359] is Compare iPhone
page: https://www.apple.com/iphone/compare/
4. click [23266] where [23266] is Pro camera system 48MP F... |
openweb_1117 | train | What are the different camera zoom levels available on the iPhone 16 Pro? | https://www.apple.com/ | https://www.apple.com/iphone-16-pro/ | 17 | keep | true | true | Trajectory navigates to iPhone 16 Pro specs and clicks camera zoom level images (24mm, 28mm, 120mm), directly addressing the objective. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [188] where [188] is
page: https://www.apple.com/store
2. click [305] where [305] is
page: https://www.apple.com/iphone/
3. click [2359] where [2359] is Compare iPhone
page: https://www.apple.com/iphone/compare/
4. click [23266] where [23266] is Pro camera system 48MP F... |
openweb_1118 | train | Compare the camera features of the iPhone 16 Pro. | https://www.apple.com/ | https://www.apple.com/iphone-16-pro/ | 21 | keep | true | true | User seeks iPhone 16 Pro camera features. Trajectory visits the iPhone 16 Pro page, navigates camera sections, and explores zoom levels and photographic styles. | claude-haiku-4-5-20251001 | page: https://www.apple.com/
1. click [188] where [188] is
page: https://www.apple.com/store
2. click [305] where [305] is
page: https://www.apple.com/iphone/
3. click [2359] where [2359] is Compare iPhone
page: https://www.apple.com/iphone/compare/
4. click [23266] where [23266] is Pro camera system 48MP F... |
openweb_112 | train | Find information about hotels near Mt. Everest. | https://www.google.com/maps | https://www.google.com/maps/search/Hotels/@27.9880925,86.8425734,12z/data=!3m1!4b1!4m7!2m6!3m5!2sMt+Everest!3s0x39e854a215bd9ebd:0x576dcf806abbab2!4m2!1d86.9249751!2d27.9881206?entry=ttu&g_ep=EgoyMDI0MTIxMS4wIKXMDSoASAFQAw%3D%3D | 3 | keep | true | true | Agent searches for Mt. Everest, then filters results to show hotels in that area. Clear objective pursued directly. | claude-haiku-4-5-20251001 | page: https://www.google.com/maps
1. type [142] [Mt. Everest ] where [142] is Search Google Maps
page: https://www.google.com/maps/place/Mt+Everest/@27.9881186,86.9043755,14z/data=!3m1!4b1!4m6!3m5!1s0x39e854a215bd9ebd:0x57…
2. click [238] where [238] is Hotels
page: https://www.google.com/maps/search/Hotels/@2... |
openweb_1121 | heldout | Find a recipe for gluten-free bread in a bread machine. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/259762/gluten-free-bread-in-a-bread-machine/ | 3 | keep | true | true | Clear instruction to find a gluten-free bread machine recipe. Trajectory directly searches for and reaches the exact recipe page requested. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [882] [gluten-free bread machine recipe ] where [882] is Search the site
page: https://www.allrecipes.com/search?q=gluten-free+bread+machine+recipe
2. goto [https://www.allrecipes.com/recipe/259762/gluten-free-bread-in-a-bread-machine/]
page: https://www.allrecipes.com... |
openweb_1122 | heldout | Leave a review for a gluten-free bread recipe. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/259762/gluten-free-bread-in-a-bread-machine/ | 9 | keep | true | true | Demonstrator found the gluten-free bread recipe and attempted to leave a review, typing review text and selecting a rating before stopping. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. click [141] where [141] is Search the site
2. type [889] [gluten-free recipes ] where [889] is Search the site
page: https://www.allrecipes.com/search?q=gluten-free+recipes
3. click [1128] where [1128] is Go to page 2 of search results
page: https://www.allrecipes.com/searc... |
openweb_1123 | heldout | Leave a review for a gluten-free recipe. | https://www.allrecipes.com/ | https://www.allrecipes.com/recipe/240496/gluten-free-sugar-cookies/ | 9 | keep | true | true | Agent searches for gluten-free recipes, navigates to a gluten-free sugar cookie recipe, writes and submits a review with rating successfully. | claude-haiku-4-5-20251001 | page: https://www.allrecipes.com/
1. type [881] [gluten-free recipes ] where [881] is Search the site
page: https://www.allrecipes.com/search?q=gluten-free+recipes
2. click [675] where [675] is A plate of decorated holiday sugar cookies in various shapes. Save Recipe CUT-OUT COOKIE RECIPES Gluten-Free Sugar Cooki... |
openweb_1124 | train | Watch a stress-relieving video from a reputable news source. | https://www.bbc.com/news | https://www.bbc.com/reel/video/p0kc9byg/a-simple-practice-that-is-good-for-stress | 5 | keep | true | true | Objective is clear and coherent. Trajectory successfully navigates BBC to find and watch a stress-relieving video, achieving the stated goal. | claude-haiku-4-5-20251001 | page: https://www.bbc.com/news
1. click [1251] where [1251] is Business
page: https://www.bbc.com/business
2. click [146] where [146] is
page: https://www.bbc.com/innovation
3. click [1280] where [1280] is A simple practice that is good for stress
page: https://www.bbc.com/reel/video/p0kc9byg/a-simple-pract... |
openweb_1125 | train | Browse through health and wellness articles | https://www.bbc.com/news | https://www.bbc.com/reel/video/p0jc5bmh/how-your-gut-microbes-shape-your-health | 9 | keep | true | true | Objective coherent: clear instruction to browse health/wellness articles. Trajectory consistent: navigates BBC health content, visits multiple wellness videos including stress, nutrition, and gut health topics. | claude-haiku-4-5-20251001 | page: https://www.bbc.com/news
1. click [1251] where [1251] is Business
page: https://www.bbc.com/business
2. click [146] where [146] is
page: https://www.bbc.com/innovation
3. click [1280] where [1280] is A simple practice that is good for stress
page: https://www.bbc.com/reel/video/p0kc9byg/a-simple-pract... |
openweb_1126 | train | Find videos and articles related to health and wellness. | https://www.bbc.com/news | https://www.bbc.com/reel/video/p0jw4jrg/how-to-reset-your-brain-with-your-breath | 13 | keep | true | true | Objective is clear and well-formed. Trajectory systematically browses BBC health and wellness videos, successfully finding multiple relevant health videos before ending at a breathing/wellness video. | claude-haiku-4-5-20251001 | page: https://www.bbc.com/news
1. click [1251] where [1251] is Business
page: https://www.bbc.com/business
2. click [146] where [146] is
page: https://www.bbc.com/innovation
3. click [1280] where [1280] is A simple practice that is good for stress
page: https://www.bbc.com/reel/video/p0kc9byg/a-simple-pract... |
openweb_1127 | train | Search for hotels in [location] on a specific date. | https://www.booking.com/ | https://www.booking.com/searchresults.html | 9 | keep | true | true | Agent searches for hotels on Booking.com with location and check-in date, successfully reaching search results page. Objective coherent and trajectory plausibly pursues it. | claude-haiku-4-5-20251001 | page: https://www.booking.com/
1. click [4186] where [4186] is Dismiss sign-in info.
2. type [329] [location] where [329] is Where are you going?
3. click [337] where [337] is Check-in Date
4. click [4333] where [4333] is 10 January 2025
5. click [340] where [340] is Check-out Date
6. click [351] where [351] is Sear... |
openweb_1128 | train | Find available check-in dates for a stay. | https://www.booking.com/ | https://www.booking.com/city/us/new-york.html | 12 | keep | true | true | Objective is clear: find available check-in dates. Trajectory demonstrates selecting check-in dates, check-out dates, and destination on Booking.com, directly pursuing the stated goal. | claude-haiku-4-5-20251001 | page: https://www.booking.com/
1. click [4187] where [4187] is Dismiss sign-in info.
2. click [330] where [330] is Where are you going?
3. click [338] where [338] is Check-in Date
4. scroll [down]
5. click [4301] where [4301] is 1 January 2025
6. click [4271] where [4271] is Next month
7. click [341] where [341] is ... |
openweb_1129 | train | Browse Data Science courses related to Finance on Coursera. | https://www.coursera.org/ | https://www.coursera.org/search?query=data%20science%20for%20finance%5D%20%5Bpress_enter_after%3D1&topic=Data%20Science&sortBy=BEST_MATCH | 5 | keep | true | true | Objective is clear and well-formed. Trajectory searches for Data Science courses related to Finance on Coursera, reaching relevant results page as intended. | claude-haiku-4-5-20251001 | page: https://www.coursera.org/
1. click [3201] where [3201] is Data Science 425 courses
2. click [3222] where [3222] is Health 471 courses
page: https://www.coursera.org/browse/data-science
3. type [2072] [data science for finance] [press_enter_after=1 ] where [2072] is What do you want to learn?
page: https:... |
End of preview. Expand in Data Studio
NNetNav-live objective filter
Per-task quality verdicts for stanfordnlp/nnetnav-live
(commit 7f69dca), produced with claude-haiku-4-5-20251001 through a distilabel
pipeline. NNetNav-live objectives were written retroactively by an LLM for exploration trajectories; some are garbled or
do not match what the recorded trajectory does. For each of the 5017 tasks the judge saw the objective, the start
and end URL and the executed actions with element names and URLs (no accessibility trees) and answered with a
structured verdict: objective_coherent, trajectory_consistent, decision (keep/drop) and a short reason.
Used to filter the training data of
johannhartmann/decider-2b-a11y-crawler-lora
(notebook 02_decision_crawler_trainieren.ipynb in the repository of the talk "Lokale Modelle selbst trainieren").
Results
| split | tasks | keep | drop |
|---|---|---|---|
| train | 4025 | 3479 | 546 |
| heldout | 992 | 866 | 126 |
| all | 5017 | 4345 | 672 |
split follows the notebook: heldout = tasks that touch amazon.com, allrecipes.com or cambridge.org (evaluation
only), train = all other tasks.
Examples of dropped objectives:
- "Check the current NBA standings." — Agent successfully navigates to NBA standings but then abandons the task by going to unrelated homepage, indicating trajectory doesn't pursue the stated objective to completion.
- "Compare Victor Wembanyama's stats to Wilt Chamberlain's stats from the 1973 season." — Trajectory finds Wilt's 1973 stats but never retrieves Wembanyama's stats for comparison as objective requires.
- "Find Java repositories on GitHub created after December 1st, 2024." — Search query contains malformed syntax ('S-01' instead of valid date format); trajectory doesn't achieve valid Java repository search after December 1st, 2024.
- "None that I can determine explicitly, but the user seems to be researching or studying the specific arXiv article, possibly preparing for the 2024-2026 Europea…" — Objective is meta-commentary about the user's possible intent rather than a clear instruction. It's a labeler's interpretation, not a user-givengoal.
- "Book a room in Las Vegas for 2 adults on January 15." — Trajectory selects January 15 as check-out, not check-in. Booking requires both check-in and check-out dates; the objective implies January 15 as check-in date.
- "Finally, since we cannot determine an objective from a sequence when there are alternatives like the objective would focus either on checking world news or lea…" — Objective is garbled meta-commentary about labeling uncertainty, not a coherent instruction. Cannot assess trajectory against incoherent objective.
Columns
task_id (NNetNav task_name), split, objective, start_url, end_url, n_steps, decision,
objective_coherent, trajectory_consistent, reason, model, prompt_trajectory (exactly the trajectory text the
judge saw). The prompt is in raw/prompt.json.
Cost and reproducibility
5065 API calls, $6.91 in total (ledger raw/api_ledger.jsonl, prices claude-haiku-4-5-20251001: $1.0/$5.0 per MTok in/out). raw/llm_cache.sqlite
holds every paid answer keyed by the exact request, so re-running the pipeline with the same prompt costs nothing.
Limitations
A single LLM judgement per task, no human review. The judge sees actions and URLs, not the pages. A dropped task is not necessarily useless, and a kept task can still contain wrong individual steps.
License and attribution
Apache-2.0, derived from NNetNav-live by Stanford NLP (Apache-2.0; Murty et al., NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild). Verdicts generated with Anthropic Claude Haiku 4.5.
- Downloads last month
- 80
Models trained or fine-tuned on johannhartmann/nnetnav-live-objective-filter