AtomicChat/DeepSeek-V4.1-Flash-NVFP4-metrics

DeepSeek-V4.1-Flash-NVFP4 metrics

Everything behind the numbers in AtomicChat/DeepSeek-V4.1-Flash-NVFP4-nvidia.

• —logprobs/lp-<run>-<corpus>.npz: the raw top-512 log probabilities of every measurement run, 49,152 scored positions each: ref, ref-repeat, ref-r3, ref-b1 (batch size 1) for the original; flat, flat-r2, flat-r3 for the uncalibrated cast; nvidia, nvidia-r2, nvidia-r3 for the calibrated checkpoint.
• —logs/kld-<run>-<corpus>.json: the KL lower bound per run against ref, coarsened on the set common to all ten runs, with per-window means and window-bootstrap intervals (nvfp4_kld.py).
• —logs/aggregate.txt and aggregate.py: per-build means over runs, run-to-run spread, paired comparisons with two intervals, bootstrap over windows only and over windows and runs.
• —logs/cast-check.py, logs/cast-check.txt: element-wise MXFP4 → NVFP4 check on 12 experts × 3 projections.
• —logs/corpora-sha256.txt: the measurement corpora as used, with a note that agentic is Muse Glimmer markup.
• —logs/coverage-*.json, amax/: calibration coverage and the raw amax dumps, enough to re-export without recalibrating.
• —logs/step-*.log, logs/progress.log: every step of the run.

Produced by atomic-quantizer/scripts/foundry-nvfp4.sh on 4×B200, 2026-09-10; the runbook explains the protocol and why the KL is a lower bound.