D_reproducibility · 完整稿件源码

% D_reproducibility.tex -- Appendix D: training, serving and release facts (added 2026-09-02 after review).
% Sources: train/scripts/runs/*.env (TP/CP/GBS/LR/MIN_LR/MAX_LENGTH), data_pipeline.md (epochs baked, pairing), bench/scripts/ops/run_eval.sh.
\section{Reproducibility}
\label{app:repro}

\paragraph{Training.}
Every student is trained with ms-swift on a Megatron backend with tensor parallelism 4, context parallelism 2 and micro-batch 1 (global batch 64), sequence length 196,608 tokens, optimizer state offloaded to CPU, on nodes of eight A100-80GB GPUs.
Sequences are pre-paired for the two context-parallel ranks (seed 42) and each epoch is an independent shuffle of the pairs, so one epoch of \vTwelve is 127 updates (8,120 rows at a global batch of 64, the last batch partial); the checkpoints were saved at updates 128 and 256, one update past the first and second epoch boundaries, and at 381, the end of training; \vThirteen runs 362 updates.
Learning-rate schedules are cosine with 5\,\% warm-up: $8\times10^{-6}\to2.4\times10^{-6}$ for the post-trained bases, $8\times10^{-6}\to8\times10^{-7}$ for \vThirteen, and $3\times10^{-5}\to3\times10^{-6}$ for the bare base; weight decay is 0.01.
Rows are rendered with the Qwen3.5 chat template; the teacher's raw thinking is kept as the thinking block (rows whose thinking survives only in summarized form are removed by the thinking-coverage rule), and loss falls on assistant text, thinking and tool calls except masked segments (\S\ref{sec:corpus}).

\paragraph{Serving and evaluation.}
Every arm runs inside \Harness 2.1.202 driven by a task runner that gives each trial a fresh container, a 262,144-token context with automatic compaction and a 5,400-second wall-time limit; students are served by SGLang behind an adapter that exposes the Anthropic Messages API and passes thinking blocks through unchanged, and the next Qwen generation is served the same way.
Sampling is at temperature 0.8, one rollout per task, and the two repeat-sample arms are second runs of the same weights; the added instruction forbids looking up the fix over the network or in the repository's git history, and network access is otherwise unrestricted.
Outcomes are the benchmark's hidden tests; a trial that times out or crashes counts as unsolved.

\paragraph{Release.}
This submission releases the fingerprint extractor, the per-arm behavioral distributions, the curation rule definitions with their thresholds and hit counts, the rule-impact analysis, the evaluation ledgers with per-task outcomes for all twelve arms, and the scripts that generate every table and figure in this paper.
Release of the corpus and the checkpoints is not part of this submission.