{ "name": "dspark-mtp-draft-head-0731reap-v1", "packaged_utc": "2026-08-12T00:39:04Z", "base_model": "0xSero/DeepSeek-V4-Flash-0731-REAP", "engine_git_rev": "85dbea6cfcd0e15651177135dc60b7b273c41445", "measurement": { "protocol": "8-prompt suite (scripts/prompt_suite.json, one per category), NGEN0>=200 so tau is past the drafter's 128-token sliding window (F92); block 6, adaptK 1.50; clean run; canonical 6-token prompt is a CONTROL, excluded from the mean (F96).", "suite_mean_tau": 3.8438, "suite_mean_tok_s": 25.5312, "base_ar_tok_s": 13.8, "n_suite_prompts": 8, "per_prompt": [ { "block": 6, "prompt": 0, "tau": 2.87, "ms_tok": 46.9, "tok_s": 21.33 }, { "block": 6, "prompt": 1, "tau": 3.64, "ms_tok": 40.7, "tok_s": 24.55 }, { "block": 6, "prompt": 2, "tau": 4.2, "ms_tok": 36.8, "tok_s": 27.16 }, { "block": 6, "prompt": 3, "tau": 2.49, "ms_tok": 53.1, "tok_s": 18.83 }, { "block": 6, "prompt": 4, "tau": 3.06, "ms_tok": 43.9, "tok_s": 22.8 }, { "block": 6, "prompt": 5, "tau": 3.94, "ms_tok": 41.0, "tok_s": 24.39 }, { "block": 6, "prompt": 6, "tau": 6.09, "ms_tok": 28.5, "tok_s": 35.08 }, { "block": 6, "prompt": 7, "tau": 3.85, "ms_tok": 37.5, "tok_s": 26.7 }, { "block": 6, "prompt": 8, "tau": 3.48, "ms_tok": 40.4, "tok_s": 24.74 } ], "lossless_gate": "PASS", "suite_tok_s": 25.5312, "suite_tau": 3.8438 }, "adaptk_at_measurement": "1.50", "source_archive": "/home/patrickd/model-backups/heads/s3", "contents": [ "config.json", "eval.log", "head_card.json", "mtp_trained.safetensors", "train_metrics.json" ], "note": "Trained MTP draft-head tensors only. The loadable head is rebuilt from a local copy of the base checkpoint with tools/build_trained_head.py, which is deterministic and refuses any tensor whose fp8 round-trip exceeds 0.10." }