Usin2705/dta_asa

DTA speaking-assessment deployment weights

Weights for the DTA Finnish speaking scorer (M-CASA). Deployment artefact, not a general model release. Loaded by the inference/ package in aalto-speech/dta-server; it will not load with plain transformers because the architecture lives in that package.

checkpoint	finnish-v3_le40_mcasa_noac_eqcap2x_ttsfluall3_s2022
test RMSE (98 held-out DTA recordings, calibrated)	0.3854
reportable CEFR range	1.14 – 3.50

Contents

scorer/model.safetensors     trained checkpoint (complete state dict)
whisper_finnish_v3/          Finnish ASR fine-tune, also the acoustic backbone. Whisper-medium.
qwen_base/                   Qwen3.5-2B base (Apache-2.0)
assets/                      task catalogue, calibrator and model card for THIS checkpoint

assets/ is uploaded alongside the weights on purpose. The container image also ships a copy, and the two must agree — dta_scorer.config.check_weights_match_assets() refuses to start if the mounted weights do not match the image's model card.

Limits

The calibrated score is clipped to [1.14, 3.50]; the model cannot certify B2 or above. Trained on Finnish L2 speech for the specific task prompts in assets/tasks.json. See assets/model_card.json for full provenance, measured metrics and known limits.