Anish13/web-agent-graph-dataset

1---2license: mit3task_categories:4- reinforcement-learning5- question-answering6language:7- en8tags:9- web-navigation10- preference-learning11- reward-modeling12size_categories:13- 1K<n<10K14---15 16# Web Agent Grouped Graph Dataset17 18This dataset contains web navigation tasks in grouped graph format with full history and candidate actions for training reward models.19 20## Dataset Description21 22- **Format**: JSON Lines (JSONL) - One entry per step with grouped candidates23- **Size**: ~2.8K step entries from 2.8K tasks24- **Domains**: GitLab, OpenStreetMap, Reddit, Shopping, Shopping Admin25 26## Data Format27 28Each line in `graph_dataset.jsonl` represents a single step with all candidate actions grouped together:29 30```json31{32  "task_id": "...",33  "goal": "Find product X and add to cart",34  "domain": "shopping",35  "step_index": 3,36  37  "history": [38    {"state_id": "S0", "screenshot": "...", "url": "...", "obs": "..."},39    {"state_id": "S1", "screenshot": "...", "url": "...", "obs": "..."},40    {"state_id": "S2", "screenshot": "...", "url": "...", "obs": "..."},41    {"state_id": "S3", "screenshot": "...", "url": "...", "obs": "..."}42  ],43  44  "current_state": {45    "state_id": "S3",46    "screenshot": "path/to/current.png",47    "url": "http://...",48    "obs": "accessibility tree..."49  },50  51  "candidates": [52    {53      "label": "gold",54      "action": "click('153')",55      "next_state": "S4",56      "next_screenshot": "path/to/next.png",57      "next_url": "http://...",58      "next_obs": "..."59    },60    {61      "label": "negative",62      "action": "click('88')",63      "negative_type": "hard_negative",64      "reason": "Leads to wrong page",65      "next_state": "S_bad1",66      "next_screenshot": "path/to/bad1.png",67      "next_url": "http://...",68      "next_obs": "..."69    }70  ]71}72```73 74## Key Features75 76- **Full History**: Complete trajectory up to current state77- **Grouped Candidates**: All possible actions (gold + negatives) from same state78- **Next-State Screenshots**: Screenshot paths for all candidate outcomes79- **Rich Metadata**: Task ID, domain, goal, step index80- **Negative Types**: Classification (easy_negative, hard_negative, detour_negative)81 82## Usage83 84```python85import json86 87# Load dataset88with open('graph_dataset.jsonl', 'r') as f:89    for line in f:90        entry = json.loads(line)91        92        # Access task info93        goal = entry['goal']94        step_idx = entry['step_index']95        96        # Access history97        history = entry['history']98        current_state = entry['current_state']99        100        # Process candidates101        for candidate in entry['candidates']:102            if candidate['label'] == 'gold':103                print(f"Gold action: {candidate['action']}")104            else:105                print(f"Negative: {candidate['action']} ({candidate['negative_type']})")106```107 108## Training Reward Models109 110This format is ideal for:1111. **Preference Learning**: Compare gold vs negative actions from same state1122. **Reward Modeling**: Predict which action leads to goal completion1133. **Action Ranking**: Rank all candidates by predicted reward114 115Example:116```python117# For each step entry118gold = [c for c in entry['candidates'] if c['label'] == 'gold'][0]119negatives = [c for c in entry['candidates'] if c['label'] == 'negative']120 121# Train model to rank gold higher than all negatives122reward_gold = model(entry['current_state'], gold)123rewards_neg = [model(entry['current_state'], neg) for neg in negatives]124```125 126## Statistics127 128See `conversion_stats.json` for detailed statistics.129 130## License131 132MIT License133