ctopd / docs /MP_OPD_INVALID_RESPONSES_VALIDATION.json
codemaivanngu's picture
fix(mp-opd): exclude unalignable responses with reason metrics
0eeed4c unverified
Raw History Blame Contribute Delete
1.49 kB
{
"date": "2026-10-19",
"platform": "WSL Linux, CPU only",
"python": "3.12.3",
"torch": "2.14.0+cpu",
"command": "TORCH_COMPILE_DISABLE=1 OMP_NUM_THREADS=1 python -m unittest discover -s tests -v",
"full_suite": {"passed": 83, "skipped": 0, "elapsed_seconds": 80.283},
"regressions": [
"Mixed replacement response exclusion with correct teacher offsets and original denominator",
"Normalization mismatch exclusion and valid-response loss/all student gradient equality across MP modes and ALIGN-NEW",
"TRUST-R/B/P replacement and normalization exclusion, B/P full-window calibration, valid-response loss/all student gradient equality",
"All-replacement microbatch connected zero loss and backward",
"Transient replacement prefix recovery remains accepted",
"Unsupported added tokens remain fatal; no Atomic fallback",
"Historical source manifests retained with explicit adapter patch hash"
],
"unchanged_donor_parity": {
"trust_max_loss_abs_error": 0.0,
"trust_max_gradient_abs_error": 5.960464477539063e-08,
"simct_max_loss_abs_error": 0.0,
"simct_max_student_gradient_abs_error": 3.725290298461914e-09
},
"not_qualified": [
"Root cause of observed company generation/retokenization mismatch",
"Frequency of excluded responses or effect on scientific comparisons",
"B200/FSDP execution of this patch or full training",
"Full training resume or unchanged Adam update when an entire window is excluded"
]
}