DeepSeek-V4.1-Flash-REAP-stats

The calibration statistics and eval data behind DeepSeek-V4.1-Flash-REAP-50 and DeepSeek-V4.1-Flash-REAP-75, pruned from deepseek-ai/DeepSeek-V4.1-Flash-1017 (revision 1bae0e108899). Independent work, not affiliated with DeepSeek.

Files

• calibration/reap-stats.npz: per-layer, per-expert REAP scores, routing counts and gate statistics over 2,946,031 calibration tokens.
• calibration/sessions/*.npz: the same statistics per calibration session, plus per-layer importance sums (24 sessions).
• survivors-{256,192,128,96}.json: the experts kept per layer for each candidate (192 is REAP-50, 96 is REAP-75).
• reap-select-*.json: selection diagnostics: overlap with frequency-based selection, never-routed experts, super-expert candidates.
• eval/sessions/*.npz: full-model (teacher) top-2,048 logit ids and values, logsumexp, tail mass and target logit at every scored eval position (108 sessions, 437,350 positions).
• student-*/: per-position KL, NLL and top-1 agreement for each candidate with its pruned experts masked.
• candidate-scores.json: those results per domain. run-config.json, progress-*.jsonl: run settings and timings.

Sources

Records came from nvidia/Open-SWE-Traces, nebius/SWE-rebench-V2-PRs, nvidia/Nemotron-Terminal-Corpus, Agent-Ark/Toucan-1.5M, CIRCL/vulnerability-scores, common-pile/stackexchange_filtered, HuggingFaceFW/finepdfs, corbt/enron-emails, allenai/WildChat-4.8M, HuggingFaceFW/finewiki. No raw text is included here, only token ids and statistics of the model's outputs on it; each source's own license and terms still apply.