babihoba's picture
Publish selected F1200, E400 comparison and original-50 teacher reproduction assets
e56d93c verified
|
Raw History Blame Contribute Delete
2.57 kB
---
license: other
license_name: bigscience-openrail-m
license_link: LICENSE
task_categories:
- text-to-speech
language:
- ko
- en
tags:
- synthetic
- distillation
- supertonic
---
# Supertonic 3 original-teacher distillation trajectories
The current F1200 reproduction data is [`round5/`](round5). These are synthetic latent tensors generated by the **original FP32 CFG4, 50-step teacher**, [Supertone/supertonic-3](https://huggingface.co/Supertone/supertonic-3), pinned revision `3cadd1ee6394adea1bd021217a0e650ede09a323`. They are not generated by a distilled student, and contain no human speech recordings.
- `round5/trajectories-original50-round5.tar.gz`:640training pairs from80unique authored sentences and128validation pairs from16unique sentences. A single voice/seed case per sentence supplies eight input/target pairs. No test trajectories were collected.
- The teacher's50-step latent states are linearly interpolated to the eight-step student grid. The final target is the exact original50-step final latent. Intermediate interpolated states are not equivalent to running the teacher for eight steps.
- `round5/extracted_manifest.json`: size/SHA256 for each of769archive members (768NPZ plus manifest). Archive members are relative paths under `trajectories-original50-round5/`, with no symlinks, owner names or machine-specific paths.
- `round5/corpus.json`: authored text used for collection. Existing evaluation sets are consumed; use a newly fixed independent set for future model selection.
- Root `asset_manifest.json`: size/SHA256 for the compressed archive and every other published artifact.
The original `trajectories-v2/` remains available for historical B reproduction:1,600training pairs from40unique sentences across five voice/seed cases, and64validation pairs. Do not mix it into the F1200 reproduction run.
Matching initialization and final weights: [F1200 model repository](https://huggingface.co/babihoba/supertonic3-optimization/tree/main/reproduction/round5). Source: [GitHub project](https://github.com/8954sood/supertonic3-optimization), [original50 collection and training record](https://github.com/8954sood/supertonic3-optimization/blob/codex/round4-evaluation-optimization/records/2026-10-29/ROUND5_LOCAL.md). Training requires more memory than the approximately1GB inference benchmark host. This small synthetic corpus does not establish broad language, speaker or perceptual quality. Applicable upstream [Open RAIL-M restrictions](LICENSE) apply; the source repository's code license does not replace them.