Dry-run staging for three proposed Hugging Face repositories: DeltaILab/Qwen3-4B-MATH-SFT-full-trajectory, DeltaILab/Qwen3-4B-MATH-SFT-step68-resume, and DeltaILab/Qwen3-4B-MATH-SFT-data-and-eval. No upload is performed.

Training used an immutable cache-backed 8753-row Arrow snapshot. The same legacy JSONL pathname later drifted to 5584 rows and is not the training input. The exported JSONL preserves Arrow order and fields. Test is the historical id_eval split under the public label test; generations are greedy with a 2048-token limit, so accuracy includes completion/truncation effects.

Step-68 restores model, ZeRO-2 optimizer shards, scheduler and consumed_samples=8704. RNG is absent, so continuation is not bitwise exact. HF snapshots are warm-start artifacts, not optimizer-state resumes. Upload scripts are dry-run by default; --execute is explicit.