Initialize sharded dataset card

README.md

@@ -1,3 +1,76 @@
 ---
-license: cc-by-nc-4.0
+license: other
+task_categories:
+  - visual-question-answering
+  - image-text-to-text
+language:
+  - en
+tags:
+  - embedded-systems
+  - multimodal
+  - procedural-reasoning
+  - benchmark
+pretty_name: Embedded-Device Tutorial Next-Step Benchmark
+size_categories:
+  - n<1K
+configs:
+  - config_name: default
+    data_files:
+      - split: test
+        path: data/test-*.parquet
 ---
+
+# Dataset Card for Embedded-Device Tutorial Next-Step Benchmark
+
+## Dataset Summary
+
+This benchmark evaluates whether multimodal models can infer the next concrete instruction in embedded-device tutorials. Each sample is a consecutive action-step pair: the anchor action provides the current visual state, and the target action provides the instruction that the model must explain.
+
+The benchmark supports two settings:
+
+- `text_only`: completed-step history plus target action summary.
+- `with_images`: the same text context plus current device photo and/or computer screenshot.
+
+## Dataset Structure
+
+The `test` split contains 216 benchmark samples from 33 selected tutorial videos. The source videos contain 249 action steps; pairing consecutive action steps within each video yields 216 samples.
+
+Columns:
+
+- `question_id`: stable sample id.
+- `device_image`: anchor device/hardware context, rendered as an HF `Image` column when available.
+- `screenshot_image`: anchor computer screenshot context, rendered as an HF `Image` column when available.
+- `history_summaries`: completed-step history, exactly the original prediction input from `target_step["step_so_far"].strip()`.
+- `task_query`: target action summary, exactly the original prediction input from `target_step["summary"].strip()`.
+- `answer`: ground-truth immediate instruction text curated in the release staging file `data/samples.jsonl`.
+- `capability`: comma-joined labels such as `hw_assembly`, `env_sw_setup`, and `code_impl`.
+- `video_id`, `anchor_step_index`, `target_step_index`: minimal provenance fields.
+
+Prompt text is not duplicated as a data column. The accompanying GitHub code constructs the exact text-only and multimodal prompts from `history_summaries` and `task_query`.
+
+In the paper prompt notation, `History Summaries` maps to `history_summaries`, `Task Query` maps to `task_query`, and visual context maps to `device_image` plus `screenshot_image` from the anchor/current-state action step.
+
+## Usage
+
+Dataset repository: https://huggingface.co/datasets/X-EASys/EmbedCopilot-Bench
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("X-EASys/EmbedCopilot-Bench")
+print(ds["test"][0])
+```
+
+## Benchmark Protocol
+
+For each video, sort steps with `is_action == "action"` by `step_index`, then pair consecutive action steps. The earlier action is the anchor/current state; the later action is the target. A model predicts the concrete next-step instruction. Evaluation compares the prediction to `answer` using the published rubric.
+
+For official GPT baselines, use the OpenAI Responses API directly with `OPENAI_API_KEY` only. Send `device_image` and `screenshot_image` as separate `input_image` payloads built from the original image bytes. Do not resize, crop, compress, collage, or convert images before API submission.
+
+## Known Issues
+
+The final 216-sample payload has no missing image references. License and redistribution terms for tutorial-derived frames must be finalized before public release.
+
+## Licensing
+
+License terms must be finalized before public release. Use `license: other` until media redistribution rights and final dataset terms are confirmed.