noteworthy changes

  • tokenizer omits <|audio|> to prevent finetunes resizing the embeddings unneccessarily
  • config.json uses the correct EOS to end generation on end of audio
  • generation_config.json uses the correct EOS to end generation on end of audio + has a reasonable default temperature
  • chat_template.jinja maps user messages to transcripts and assistant messages to audio for simple conversational context management
Downloads last month
180
Safetensors
Model size
3.3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for nytopop/3b_or_base

Finetuned
(72)
this model