ctopd / docs /TRUST_PAPER_VALIDATION.json
codemaivanngu's picture
feat(kd): add TRUST R B P and separate SimCT paper scores
619cff6 unverified
Raw History Blame Contribute Delete
2.29 kB
{
"date": "2026-10-19",
"donor_commit": "529ee4b691732e1040d6a45f05fd12f58def96ec",
"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": 76, "skipped": 0, "elapsed_seconds": 77.342},
"trust_parity": {
"cases": 6,
"max_loss_abs_error": 0.0,
"max_gradient_abs_error": 5.960464477539063e-08,
"dtype": "float32",
"atol": 1e-6,
"rtol": 1e-5,
"coverage": "TRUST-R/B/P complete loss and all student gradients, with/without final softcap",
"oracle": "Original donor TRUST kernels in separate fixture namespace and original TRUST-R method body; frozen source-verified Gram/credit helpers shared"
},
"paper_score": {
"result": "pass",
"scope": "Eq7-8 score reconstruction on released alignment/candidate interface",
"checks": ["Independent candidate scores and RKL", "Complete adapter loss and all student gradients", "Per-position additive logit-shift loss/gradient invariance"],
"loss_gradient_atol": 1e-6,
"loss_gradient_rtol": 1e-5,
"shift_invariance_atol": 2e-6
},
"additional_checks": [
"Explicit-autograd full-batch cross-response geometry and TRUST-P opposing-component projection",
"Actual actor fit/evaluate dispatch with real CPU SGD: one optimizer update, train-mode dropout replay, Python/NumPy/Torch RNG advance once",
"Microbatch regrouping preserves full-window loss and gradients",
"Evaluation restores model mode and leaves weights/calibration state untouched",
"Common-normalizer/window guards and calibration cleanup on errors",
"No teacher parameter or hidden-state gradient",
"All existing Strict/SimCT/MP/ALIGN numerical regressions and frozen source hashes",
"Launcher preserves common model/data/batch/steps/evaluation settings",
"git diff --check"
],
"not_qualified": [
"Target Torch 2.11 CUDA, B200/FSDP runtime and full-size memory/throughput",
"Full Transformer or Adam descent guarantees",
"Full training resume, accuracy or either paper's full recipe",
"Complete SimCT paper semantics beyond Eq7-8 score correction",
"Parity with unavailable Rethinking simct baseline revision"
]
}