KIT_IWSLT26_ModelCompression_en-de_unconstrained_primary

 - Exact storage size: 3.98GB
 - Number of parameters: 8.2B
 - Applied Compression techniques: HQQ 4 bit quantization on all linear layers, 2 bit quantization on word embeddings, 3-bit quantization of MLP in 4 selected layers, AWQ calibration on ACL6060 dev set.
 - Datasets used: The base model was finetuned with LoRA adapters using a 200k-sentence subset of the Europarl and CoVoST corpora. The fine-tuning data covers English-to-X translation across six target languages: DE, ZH, ES, NL, PT, and TR 
 - Any relevant additional details: The difference to system "contrastive2" is applying repetition penalty (for TVSERIES using pyannote instead of SHAS).
From\To de zh
en N/A N/A