README.md

metadata

license: mit
task_categories:
  - other
tags:
  - web-trajectory
  - information-seeking
  - user-behavior
  - trial-and-error
  - problem-solving
size_categories:
  - 1K<n<10K

TEC: A Collection of Human Trial-and-error Trajectories for Problem Solving

This is the dataset for the paper:

TEC: A Collection of Human Trial-and-error Trajectories for Problem Solving Xinkai Zhang, Jingtao Zhan, Yiqun Liu, and Qingyao Ai.

Dataset Description

TEC captures the human trial-and-error process in web search with both behavioral traces and structured diagnostic reflections. Each record represents a multi-trial task trajectory where participants iteratively search, attempt answers, reflect on failures, and retry with corrective plans.

Each task record includes:

• Task information: Open-domain factoid question, ground truth answer, completion status
• Participant profile: Demographics and expertise levels (anonymized)
• Pre-task annotation: Familiarity, difficulty prediction, initial search query, initial guess
• Trial outcomes: Per-trial answers with correctness labels, confidence, and formulation method
• Evidence markers: Selected text, DOM position, source URL with relevance/credibility ratings
• Reflection annotations (on failure): Prioritized failure diagnosis, corrective plan, adjusted difficulty
• Post-task annotation (on success): Actual difficulty, "aha" moments, unhelpful paths, strategy shifts
• Cancellation annotation (on giving up): Cancellation reason, missing resources
• Behavioral traces: Full rrweb DOM recordings, interaction events, and mouse movements per page

Dataset Statistics

Metric	Count
Participants	46
Tasks	2424
Trials	5370
Webpages	41229

Anonymization

This dataset has been anonymized:

• User identifiers (username, email, name, phone) are replaced with [ANONYMIZED]
• Participant IDs are replaced with sequential identifiers (e.g., participant_000001)
• Age is binned into ranges (e.g., 25-34)
• Profile images and field of expertise are anonymized

Data Format

The dataset is provided in Parquet format. Each row is a complete task record.

Behavioral trace fields (rrweb_record, event_list, mouse_moves, page_switch_record) are stored as JSON strings due to their variable nested structure. All other fields are native types.

Schema

{
  "task_id": 1,
  "participant_id": "participant_000001",
  "question": "What is ...",
  "ground_truth": "...",
  "status": "completed",
  "start_timestamp": "2024-01-15T10:30:00Z",
  "end_timestamp": "2024-01-15T10:45:00Z",
  "num_trial": 2,
  "participant": {
    "username": "[ANONYMIZED]",
    "profile": {
      "age": "25-34",
      "gender": "M",
      "occupation": "researcher",
      "education": "phd"
    }
  },
  "pre_task_annotation": {
    "familiarity": 2,
    "difficulty": 1,
    "first_search_query": "...",
    "initial_guess": "...",
    "expected_source": ["search_engine"]
  },
  "post_task_annotation": {
    "difficulty_actual": 3,
    "aha_moment_type": "search_result",
    "strategy_shift": ["..."],
    "strategy_shift_other": "",
    "unhelpful_paths": ["..."]
  },
  "cancel_annotation": {
    "category": [],
    "reason": "",
    "missing_resource": ""
  },
  "trials": [
    {
      "trial_num": 1,
      "answer": "...",
      "is_correct": false,
      "confidence": 3,
      "reflection_annotation": {
        "failure_category": "Ineffective Search",
        "corrective_plan": "Improve Search",
        "adjusted_difficulty": 4,
        "notes": "..."
      },
      "justifications": [
        {
          "url": "https://...",
          "text": "selected text",
          "dom_position": "CSS selector",
          "relevance": 0.8,
          "credibility": 0.9
        }
      ],
      "webpages": [
        {
          "title": "Page Title",
          "url": "https://...",
          "referrer": "https://...",
          "dwell_time": 45,
          "rrweb_record": "[{...}]",
          "event_list": "[{...}]",
          "mouse_moves": "[{...}]",
          "page_switch_record": "[{...}]"
        }
      ]
    }
  ]
}

Key Fields

Record	Key Fields
Webpage (per page)	URL, title, rrweb DOM recording, interaction events, mouse/scroll trajectory, dwell time, referrer
Trial outcome (per trial)	Answer, correctness, confidence, formulation method
Evidence	Selected text, DOM position, source URL, relevance/credibility ratings
Reflection (on failure)	Failure diagnosis (prioritized), corrective plan (prioritized), adjusted difficulty
Pre-task (per task)	Familiarity, difficulty estimate, initial query plan, initial guess
Post-task	Actual difficulty, "aha" moment type, unhelpful paths, strategy shifts
Cancellation	Cancellation reason, missing resources

Usage

from datasets import load_dataset

dataset = load_dataset("Serendipity2004/TEC", split="train")

# Access a task trajectory — all fields are native dicts/lists
task = dataset[0]
print(task["question"])
print(f"Number of trials: {task['num_trial']}")
print(f"Participant age: {task['participant']['profile']['age']}")

# Iterate over trials (native dicts, no json.loads needed)
for trial in task["trials"]:
    print(f"Trial {trial['trial_num']}: correct={trial['is_correct']}")
    if trial["reflection_annotation"]["failure_category"]:
        print(f"  Failure: {trial['reflection_annotation']['failure_category']}")

# Large behavioral data fields are JSON strings — parse when needed
import json
for trial in task["trials"]:
    for wp in trial["webpages"]:
        events = json.loads(wp["event_list"]) if wp["event_list"] else []
        print(f"  Page: {wp['url']} ({len(events)} events)")

Citation

@article{zhang2026tec,
  title={TEC: A Collection of Human Trial-and-error Trajectories for Problem Solving},
  author={Zhang, Xinkai and Zhan, Jingtao and Liu, Yiqun and Ai, Qingyao},
  year={2026}
}

License

MIT License

Exported

• Date: 2026-02-13T05:45:43.586770+00:00
• Anonymized: Yes