You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

NB-ASR UNN NorMed 100% 10k

Access: This is an internal test model. Please request access to use its files.

Internal synthetic demo for evaluation. This model is part of the NB-ASR project and NorMed.

The experiment compares a Norwegian NB-ASR control with mixtures containing 10%, 50%, or 100% normalized synthetic NorMed training examples. All models were fine-tuned for 10,000 optimizer steps from the same NB-ASR Qwen3-ASR 0.6B checkpoint.

Model NorMed synthetic mix Rows (NB / NorMed) No-prompt reading-optimised WER CER No-prompt overall WER
NB control 0% 205,556 / 0 18.20% 7.51% 18.78%
NorMed 10 10% 925,002 / 102,778 17.90% 7.34% 18.51%
NorMed 50 50% 102,778 / 102,778 18.35% 7.44% 18.90%
NorMed 100 100% 0 / 102,778 18.61% 7.20% 18.95%

Scores are from the shared Sprakradet evaluation at step 10,000. The 0%, 10%, and 50% manifests use an empty prompt for normalized examples; the 100% run used the earlier frozen manifest. These are experimental internal checkpoints, not production releases.

Downloads last month
10
Safetensors
Model size
0.8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including NbAiLab/nb-asr-unn-normed100-10k