voxscore model bundle
Pinned mirror of the models used by voxscore, a self-hosted scoring pipeline for spoken open-ended assessment responses. Nothing here is original work — these are redistributions at fixed revisions so that a scoring run is reproducible and can be performed on a machine with limited network access.
Contents
| Source | Licence | Revision | Size | Used for |
|---|---|---|---|---|
openai/whisper-large-v3 |
Apache-2.0 | 06f233fe06e7 |
3.09 GB | Transcription + per-window language identification |
facebook/wav2vec2-base-960h |
Apache-2.0 | 22aad52d435e |
0.38 GB | CTC forced alignment for word-level timings |
BAAI/bge-m3 |
MIT | 5617a9f61b02 |
2.29 GB | Sentence embeddings for relevance and repetition |
MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli |
MIT | 6f5cf0a2b59c |
0.38 GB | Entailment coverage and stance detection |
Unbabel/gec-t5_small |
Apache-2.0 | c958d53bfbce |
0.24 GB | Grammatical error correction, diffed by ERRANT into typed errors |
Every model is Apache-2.0 or MIT, both of which permit redistribution with attribution. Full credit belongs to the original authors; please cite and follow the licence of the upstream repository rather than this mirror.
.bin / .h5 duplicates are omitted where safetensors are available.
Use
from huggingface_hub import snapshot_download
path = snapshot_download("Pransfrance/voxscore-models", repo_type="model")
# then point voxscore at it:
# VOXSCORE_MODEL_DIR=<path>
Or fetch a single model:
snapshot_download("Pransfrance/voxscore-models", allow_patterns=["openai__whisper-large-v3/*"])
manifest.json carries the source repo, revision and licence for each entry in
machine-readable form.
Why a mirror
Pinning (upstream repos change and are occasionally removed), one download instead of five, and the ability to move a single directory to a restricted machine. If you can reach the Hub, prefer the original repositories.