Add README

README.md

@@ -1,15 +1,43 @@
----
-dataset_info:
-  features: []
-  splits:
-  - name: train
-    num_bytes: 0
-    num_examples: 0
-  download_size: 423
-  dataset_size: 0
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
----
+# Web Agent Graph Dataset
+
+This dataset contains web navigation tasks in graph format with single-step negatives.
+
+## Format
+
+Each line in `graph_dataset.jsonl` is a JSON object with:
+
+- `graph_data`: Task metadata (domain, root_url, goal)
+- `node_data`: Current state information (url, observation, screenshot path)
+- `step_data`: Action and outcome
+  - `action`: The action taken
+  - `obs`: Next state observation (axtree_txt, url)
+  - `step_number`: Step index in the trajectory
+  - `is_positive`: Boolean indicating if this is a positive (gold) or negative action
+  - `negative_type`: (Only for negatives) Type of negative: "easy_negative", "hard_negative", or "detour_negative"
+  - `negative_reason`: (Only for negatives) Explanation for why this action is negative
+
+## Usage
+
+```python
+import json
+
+# Load dataset
+with open('graph_dataset.jsonl', 'r') as f:
+    for line in f:
+        entry = json.loads(line)
+        # Process entry
+        print(entry['graph_data']['goal'])
+        print(entry['step_data']['action'])
+        print(entry['step_data']['is_positive'])
+```
+
+## Statistics
+
+See `conversion_stats.json` for dataset statistics including:
+- Total tasks and entries
+- Positive vs negative distribution
+- Domain breakdown
+
+## Citation
+
+If you use this dataset, please cite the original work.