CSY20060325/sjj

Synthetic training data, four evaluation sets, and V3 L1 best weights

This release contains only selected synthetic training samples, the frozen L1/L2 full/hard evaluation sets, and the full merged V3 L1 calibrated250 checkpoint. The publisher has confirmed permission to redistribute the included data. Original source rights and attribution remain applicable; no blanket replacement license is asserted.

Layout

• releases/20261004/train/: independently extractable shards, each with manifest.jsonl, selected audio, source hashes, and SHA256SUMS.json. Union membership is specified by selected_for_l1 and selected_for_l2; do not count a shared sample twice. No real training samples, rejected candidate regions, failed generations, or feature caches are included.
• releases/20261004/tests/l1_full2232.tar: L1 full, 2,232 unique segments.
• releases/20261004/tests/l1_hard553.tar: L1 hard, 553 unique segments.
• releases/20261004/tests/l2_full1596.tar: L2 full, 1,596 unique segments.
• releases/20261004/tests/l2_hard407.tar: L2 hard, 407 unique segments.
• releases/20261004/weights/v3_l1_best_calibrated250.tar: unchanged merged full checkpoint and provenance. This is not the L2 best or Stage3 checkpoint.
• releases/20261004/release_index.json: final artifact sizes, SHA256 hashes and exact counts. The release is complete only when this index is present and every listed artifact exists.

Training selection: L1 830,976 unique synthetic segments; L2 388,004 unique synthetic segments. Repeated concat4 tail-padding instances are not duplicated. Fine labels in the L1 training pool are retained as source annotations, not represented as independently validated L2 labels; use L2 membership for the selected L2 training subset. Historical disgust/aversion selection is preserved, not revised.

Base synthetic audio uses Qwen3-TTS, MOSS-VoiceGenerator and VoxCPM2; a separately selected SeedAudio supplement is included. Source text was generated with Qwen3-4B. Batch names and source provenance are retained in manifests. WAV sources are cropped to the actual selected training intervals with the original rounded native-frame boundaries and saved losslessly without resampling; sample values are checked for equality after serialization. Full-utterance SeedAudio MP3 bytes and original crop bounds are preserved to avoid changing lossy decoder behavior. Source-relative paths are provenance only; use the packaged audio paths.

Each test archive contains the original single-segment audio and original frozen concat4 audio. members.jsonl supplies references; concat4.jsonl specifies interval boundaries, member UID and score mask. Never count padding where score=false. L1 archives expose L1 reference labels only; L2 archives expose both levels. These four datasets overlap, and hard sets are subsets rather than independent corpora. They were repeatedly used for development/model selection; do not describe them as independent unseen test sets or claim unseen-set SOTA.

Evaluation data include real speech and synthetic speech. Source identifiers, domains and archive source-manifest hashes are preserved; provenance paths can identify upstream corpora (including M3ED and EARS). Redistribution permission has been confirmed by the uploader, not independently adjudicated here. Model implementation: https://github.com/uuuhh/emo . The checkpoint contains model parameters/config; any base-model/code dependencies retain their original terms.

Extract each tar into a separate directory: audio paths in its manifest are relative to that directory. Do not extract unrelated shard manifests into one common directory, because manifest filenames repeat. Check every downloaded tar against the final release index, then the extracted files against its SHA256SUMS.json.