APG_IWSLT26_ModelCompression_en-zh_constrained_primary
Primary constrained submission for IWSLT 2026 Model Compression, English to Chinese.
Base model: Qwen/Qwen2-Audio-7B-Instruct
Compression method: selected-layer codec compression
Setting: q3 + zstd-3
Layer subset: all_selected
Selected layer count: 96
Applied layer count: 96
Compressed layers: transformer MLP layers, gate_proj, up_proj, down_proj, across language_model.model.layers.0 to 31
Artifact size: 14527608574 bytes, 13.5299 GB
Track condition: constrained
Predictions prepared as plain UTF-8 text, one sentence per line.
Base model: Qwen/Qwen2-Audio-7B-Instruct
Compression method: selected-layer codec compression
Setting: q3 + zstd-3
Layer subset: all_selected
Selected layer count: 96
Applied layer count: 96
Compressed layers: transformer MLP layers, gate_proj, up_proj, down_proj, across language_model.model.layers.0 to 31
Artifact size: 14527608574 bytes, 13.5299 GB
Track condition: constrained
Predictions prepared as plain UTF-8 text, one sentence per line.
| From\To | de | zh |
|---|---|---|
| en | N/A | N/A |