dyyfk/minicpm-o45-native-gate-data

Official-native gate training/eval bundle for issue #8 (deployed gate_native.json, 8bq recipe, train_n=5228).

Contents (official_native_bundle.zip, 187 files, unpacks to official_native_bundle/data/):

• —Training tags caliboff, expoff, exp2off, exp3off, exp3zhoff, freshoff: frozen_native_<tag>_feats.shard*.npz (ids + 12288-d float32 X), frozen_native_<tag>_traces.jsonl.shard* (official-native answer_text, no_speak, eot_seen, n_ans_chunks), frozen_native_<tag>_judged.parquet (id, query, referenceanswer, answer, adequate, judgereason, escalate_label).
• —Eval tags testoff, striviaqaoff, swebqoff, sllamaoff, sdqaoff, sreasonoff: same three file kinds.
• —Label/metadata files scripts/31 reads: calib_features.parquet, expansion{,2,3,3zh}_labels.parquet, fresh_labels.parquet, frozen_v3_traces.parquet; query files queries*.jsonl for every pool above; the deployed gate_native.json.
• —MANIFEST.json: per-tag row counts / dim / shard paths, per-file sha256, serving-config identifier (official_cfg: topk=20, forcelisten_count=3, system prompt "You are a friendly assistant."), source commit. SHA256SUMS: sha256sum -c format.

Verified from a clean extraction: sha256sum -c SHA256SUMS passes; scripts/31_official_refit_labels.py --source native --dry reconstructs train 5228 (core 4986 + fresh 242), OOF AUC 0.848, thresholds identical to the deployed artifact.

Built at commit 03bf0e4 by scripts/32_export_official_bundle.py. No credentials included.