Стенд-проверка (stage-2) · промт оператора:

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "The congestion-aware router was implemented and evaluated end-to-end on the registered custom multi-token MoE track. It reduced trained-model expert-load CV substantially, confirming the proposed mechanism, but test MSE was worse than the tuned independent-softmax baseline and the paired permutation test was not significant.", "metrics": { "baseline": "MSE mean 0.9469610155, std 0.2249618225; best lr=0.006, temperature=0.5.", "idea": "MSE mean 0.9704978541, std 0.2190361309; best lr=0.006, temperature=0.7. Load-CV delta=-0.2865399508." }, "how_to_run": "CUDA_VISIBLE_DEVICES='' /home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "custom_moe_track.py", "bench_report.json" ], "limitations": "Small custom registered track, tiny four-expert model, 240 training examples, 100 test examples, four epochs, and CPU fallback due to host CUDA initialization failure. Production-scale Transformer/MoE, hard top-k dispatch, overflow/drop rate, throughput, EMA capacity, straight-through routing, and large-scale expert-FLOP parity were not tested.", "bench_report": { "bench_version": 1, "track": "congestion_moe_sequence", "model": "tiny_moe", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "temperature": 0.5 }, "sweep": [ { "cfg": { "lr": 0.0015, "temperature": 0.5 }, "mean": 1.013328194618225 }, { "cfg": { "lr": 0.0015, "temperature": 0.7 }, "mean": 1.0138949006795883 }, { "cfg": { "lr": 0.0015, "temperature": 1.0 }, "mean": 1.0139495730400085 }, { "cfg": { "lr": 0.003, "temperature": 0.5 }, "mean": 0.9528659582138062 }, { "cfg": { "lr": 0.003, "temperature": 0.7 }, "mean": 0.9548223316669464 }, { "cfg": { "lr": 0.003, "temperature": 1.0 }, "mean": 0.9556002467870712 }, { "cfg": { "lr": 0.006, "temperature": 0.5 }, "mean": 0.853488102555275 }, { "cfg": { "lr": 0.006, "temperature": 0.7 }, "mean": 0.859209343791008 }, { "cfg": { "lr": 0.006, "temperature": 1.0 }, "mean": 0.8618226200342178 } ], "full": { "mean": 0.9469610154628754, "std": 0.22496182254054511, "per_seed": [ 0.7249739170074463, 0.5741610527038574, 0.8283010125160217, 1.2865164279937744, 1.0606132745742798, 1.210940957069397, 0.897042989730835, 0.9931384921073914 ], "n": 8 } }, "idea": { "mean": 0.9704978540539742, "std": 0.21903613092802948, "per_seed": [ 0.7384425401687622, 0.5904383659362793, 0.8622050285339355, 1.2669295072555542, 1.1580994129180908, 1.1912760734558105, 0.947786271572113, 1.0088056325912476 ], "n": 8 }, "comparison": { "delta_mean": 0.023536838591098785, "idea_wins": 2, "n_pairs": 8, "per_seed_diffs": [ 0.013468623161315918, 0.016277313232421875, 0.03390401601791382, -0.019586920738220215, 0.09748613834381104, -0.019664883613586426, 0.050743281841278076, 0.0156671404838562 ], "p_value": 0.13795, "mde": 0.03200907134547522, "mde_rel_pct": 3.380188922542831, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "prediction": "congestion-aware routing reduces expert-load CV while retaining a fixed row budget", "predicted_direction": "negative", "observed_load_cv_delta": -0.28653995075335703, "confirmed": true }, "custom_track": { "name": "congestion_moe_sequence", "file": "custom_moe_track.py", "domain": "moe-routing" } }, "system_verdict": "partial", "practical_verdict": "inconclusive", "mechanism_ok": 1, "system_judged": true }