ctopd / docs /SIMCT_VALIDATION.json
codemaivanngu's picture
feat(simct): add isolated author baseline with loss gradient parity
43eefa6 unverified
Raw History Blame Contribute Delete
1.87 kB
{
"date": "2026-10-19",
"source_commit": "cf0f33a0e6c967d4b74ea32b2dba12be01b73b9e",
"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": 56,
"skipped": 0,
"elapsed_seconds": 91.714
},
"original_source_parity": {
"cases": 6,
"max_loss_abs_error": 0.0,
"max_student_gradient_abs_error": 3.725290298461914e-09,
"atol": 1e-6,
"rtol": 1e-5,
"dtype": "float32",
"coverage": "Mixed spans with two responses, strict-only, student softcap, G(h) disabled and three positive thresholds",
"oracle": "Pinned original class and original RKL numerical bodies; only imports/decorators removed for isolated execution"
},
"additional_checks": [
"Independent hand-constructed virtual logits, RKL and gradients",
"No teacher parameter or hidden-state gradient",
"Response boundaries and shared microbatch normalization",
"Empty alignment, invalid hidden rows, arguments and topology guards",
"Existing Strict/MP source hashes and numerical regression tests",
"SimCT dry-run retains batch512, steps100, generation8192 and common optimizer/evaluation settings",
"Staged diff whitespace check excluding the verbatim original RKL fixture, which retains three upstream trailing-whitespace lines",
"Source hashes rechecked after renaming the MIT notice to enforce LF through existing .gitattributes"
],
"not_qualified": [
"Paper-equation equivalence: released code averages raw logits rather than log-probabilities",
"Parity with other KDFlow implementations named simct",
"Target Torch 2.11 CUDA, B200/FSDP execution and full-size memory/throughput",
"Full training, published scores and the original SFT warmup recipe"
]
}