plnguyen2908/sam-audio-v1

SAM-Audio v1: joint Qwen/SAM fine-tuning weights

Final live weights at step 55,534, eight epochs (888,544 target/prompt examples processed). Training data: URMP, ChoralSynth, MSDWild and EasyCom; 111,068 training examples.

This is a custom joint-model delta, not a standalone original SAM-Audio checkpoint or a Transformers from_pretrained model. It requires the matching joint Qwen/bridge/SAM architecture, LoRA insertion and the original frozen base weights. No base-model weights are included in trained_parameters.safetensors. Runtime buffers are supplied separately.

Files and loading

• trained_parameters.safetensors: all learned named parameters (Qwen adapters, bridge, SAM LoRA and trainable SAM conditions).
• model_buffers.safetensors: checkpointed named buffers.
• training_config.json: exact effective configuration and named parameter groups.
• sam_tuning.json: SAM tuning settings.

After constructing the matching joint module (the unwrapped module, without a module. prefix), apply the tensors as follows. This does not construct that architecture by itself:

import torch
from safetensors.torch import load_file

parameters = dict(joint.named_parameters())
buffers = dict(joint.named_buffers())
with torch.no_grad():
    for filename, destination in [
        ("trained_parameters.safetensors", parameters),
        ("model_buffers.safetensors", buffers),
    ]:
        for name, value in load_file(filename).items():
            assert name in destination, name
            assert destination[name].shape == value.shape, name
            destination[name].copy_(value)
joint.eval()

Evaluation

817 held-out target/prompt examples, live weights, midpoint solver, 16 steps. Final example-weighted average SI-SDR improvement over the input mixture: +2.8064 dB. URMP +3.0443; ChoralSynth +9.1767; MSDWild -1.4566; EasyCom +0.2462 dB. These values are not an improvement relative to unmodified SAM-Audio. See validation_summary.csv.

The SAM License is reproduced in LICENSE. Applicable base-model licenses and access conditions still apply. Optimizer, RNG and rank-local training-resume state are not included in this inference export.