viq / surgery_report.json
amir0907's picture
Upload folder using huggingface_hub
ba4245a verified
Raw
History Blame Contribute Delete
901 Bytes
{
"source_old_model": "microsoft/VibeVoice-1.5B",
"source_qwen3_model": "Qwen/Qwen3-4B-Instruct-2507",
"old_decoder_hidden_size": 1536,
"new_decoder_hidden_size": 2560,
"new_decoder_num_layers": 36,
"new_decoder_num_heads": 32,
"new_decoder_num_kv_heads": 8,
"new_diffusion_hidden_size": 2560,
"surgery_strategy": "function-preserving zero-padding (Net2Net-style)",
"new_max_position_embeddings": 65536,
"notes": "The model has been surgically expanded. The LLM weights are fully trained (loaded from Qwen3-4B-Instruct-2507). The non-LLM weights (tokenizers, connectors, diffusion head) have been expanded with function-preserving zero-padding so that their behaviour on the OLD 1536-D sub-space is preserved. The new 1024 dims start at zero / near-zero and will be learned during fine-tuning. A short fine-tuning run is REQUIRED before the model produces high-quality speech."
}