onereason_221k_sft_exactdedup

# OneReason 221K SFT Exact-Dedup Dataset Strict paper-ratio SFT data built for the OneReason post-training experiments. Files: - `stage3_r3anchored_exactdedup_train.jsonl`: 216,854 train samples. - `stage3_r3anchored_exactdedup_valid.jsonl`: 4,427 validation samples. - `stage3_r3anchored_exactdedup_manifest.json`: data ratios, counts, and duplicate verification. Total samples: 221,281. Exact duplicate extra count: 0. Paper-ratio targets: - R0 perception: 23.82% - R1 derivation: 10.12% - R2 evolution: 3.13% - R3 recommendation: 22.36% - Itemic instruction: 2.61% - General-domain replay: 37.96%