EdgeBench-Home

EdgeBench-Home is one bilingual smart-home agent benchmark with seven task families and two evaluation views.

Dataset size

• 420 Chinese canonical cases
• 420 English canonical cases
• 840 canonical bilingual cases in total
• Each canonical case has both a State-Given (SG) and a State-Seeking (SS) view; these are derived views, not additional cases.

Each language contains 60 cases in each task family. Chinese and English records are aligned by case ID and reported as separate language tracks.

Task families

Abbreviation	Task family
DGC	Device Grounding and Control
MSC	Multi-device Selection and Control
AIR	Ambiguity and Intent Resolution
FAR	Future Automation Registration
DCR	Dialogue-context Resolution
PMU	Personalized Memory Use
REQ	Read-only Environment Queries

Layout

EdgeBench-Home/
  zh/ and en/
    DGC/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    MSC/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    AIR/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    FAR/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    DCR/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    PMU/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    REQ/data/{cases.jsonl,evaluation.jsonl,device_specs.json}
    State-Given/data/{DGC,MSC,AIR,FAR,DCR,PMU,REQ}.jsonl
    State-Given/protocols/{DGC,MSC,AIR,FAR,DCR,PMU,REQ}.json
    State-Seeking/data/{DGC,MSC,AIR,FAR,DCR,PMU,REQ}.jsonl
    State-Seeking/protocols/{DGC,MSC,AIR,FAR,DCR,PMU,REQ}.json
  viewer/{zh,en}/{state_given,state_seeking,cases,evaluation}.parquet

The four *_state_given and *_state_seeking configurations are model-input views. They use HF-friendly Parquet files with stable columns; nested inputs are stored in model_input_json and can be recovered with json.loads. The raw JSONL views under State-Given/ and State-Seeking/ retain native JSON objects for the evaluation runner. The *_cases and *_evaluation configurations expose the canonical cases and evaluator-only annotations separately; their nested fields are likewise serialized as JSON strings in the Parquet view.

Using the views

For State-Given inference, send only model_input from the corresponding file under <language>/State-Given/data/.

For State-Seeking inference, send only model_input from the corresponding file under <language>/State-Seeking/data/. The evaluation runtime uses the matching canonical case, device specifications, and protocol file while executing tool calls.

Never provide evaluation.jsonl, outer record IDs, manifests, or validation metadata to the model. These contain evaluator-only information used for scoring. For inference from the raw JSONL views, send only each input row's model_input value. For an HF configuration, parse model_input_json and send the resulting object.

Reporting

Report Chinese and English predictions and scores separately. Do not combine the language tracks into one score without also reporting both tracks independently.

The data package contains frozen data and protocols. The evaluation runtime is maintained separately.

See DATASET_MANIFEST.json and VALIDATION_REPORT.json for file hashes and release validation.