Quantization made by Richard Erkhov.
Fintuned Llama for our music generation task.
The tokenization of the model was not touched.
The model has three values, instruction, input and output.
Set the instruction to the simplified version of the “attributes” tag. This contains data like the tempo, and key that is kept throughout the entire length of the song.
The tokens in the instruction input are kept as important for the attention mechanism.
Finally, the existing portion of the song is put into the input, and the model then generates future tokens in the output
This llama model was trained 2x faster with
Unsloth and Huggingface's TRL library.
