sshchoholiev/research-paper-read-bench

Research Paper Read - On-device Benchmark Submissions

Append-only collection of benchmark results submitted from the Research Paper Read iOS app.

Each submission is a JSON file under submissions/{submission_id}.json written by the leaderboard Space after validation, idempotency check, and rate limiting.

See schema.json for the full field shape. Headlines:

• —prompt_tps / gen_tps — measured tokens/sec for prefill and 256-token greedy decode
• —device_id / device_name — sysctl hw.machine + display-name mapping
• —model_name / model_quant — Qwen3.5-4B / Q4_K_M at the time of writing
• —thermal_state — nominal | fair | serious | critical (ProcessInfo.thermalState)
• —bench_version — bump when the bundled prompt or token budget changes

Filtering convention used by the Space's public leaderboard:

bench_version == 1
thermal_state == "nominal"
is_simulator == false

then group by device_id, take max gen_tps.