metadata
license: apache-2.0
language:
- en
This is a MoE-ification of TinyLlama/TinyLlama-1.1B-Chat-v1.0 using the Mixtral branch of mergekit
The Goal was to MoE-fy the TinyLlama model and then use this as a base model to finetune from. The intuition being finetuning 8x1b should give better performance than finetuning 1b by itself.
More work coming!
Chat Template
def make_prompt(instruction):
return f"<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n"
llm.generate(make_prompt('What is quantum tunneling?'))
Mergekit Config
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
gate_mode: hidden
dtype: bfloat16
experts:
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]
- source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
positive_prompts: [""]