| [13:13:50] phase 5a: s2d = s2 + 4 epochs on full_dual |
| [13:19:55] s2d raw val300: MICRO-F1 0.9990 | P 1.0000 R 0.9979 | tp/fp/fn 480/0/1 | m | chat val300: MICRO-F1 0.9979 | P 1.0000 R 0.9958 | tp/fp/fn 479/0/2 | m |
| [13:21:50] s2d onnx_rr: MICRO-F1 0.9986 malformed 0/200 |
| [13:22:18] s2d calib: {"artifact": "onnx_s2d/final_rr", "ttft_p50": 74, "ttft_p95": 75, "tok_s": 92.06, "cold_ms": 2208, "rss_MiB": 539, "size_MiB": 339, "requests": 200, "failures": |
| [13:22:32] staged onnx_s2d_rr |
| [13:22:39] gguf s2d converted |
| [13:23:20] === gguf_s2d first 100: MICRO-F1 1.0000 malformed 0 | task ttft p50 75 ms probe 0: ok=True ttft=0.0 ms total=142 tokens=0 output='' probe 1: ok=True ttft=0.0 ms total=162 tokens=0 output='' probe 2: ok=True ttft=0.0 ms total=166 tokens=0 output='' |
| [13:23:55] gguf s2d calib: {"artifact": "gguf_s2d", "ttft_p50": 125, "ttft_p95": 126, "tok_s": 85.85, "cold_ms": 559, "rss_MiB": 430, "size_MiB": 195, "requests": 200, "failures": 0, "wall_s": 35.2} |
| [13:24:06] staged gguf_s2d |
| [13:24:07] phase 5b: deep extract slate rebuilt with reduce_range |
| [13:27:11] r12c onnx_rr: MICRO-F1 0.9615 malformed 0/150 |
| [13:27:25] staged onnx_r12c_rr |
| [13:30:59] b16v2 onnx_rr: MICRO-F1 0.9859 malformed 0/150 |
| [13:31:15] staged onnx_b16v2_rr |
| [13:36:36] c28 onnx_rr: MICRO-F1 0.9894 malformed 0/150 |
| [13:36:52] staged onnx_c28_rr |
| [13:36:52] PHASE5_DONE |
|
|
