🦙⛰️ BigLlama-3.1-681B-Instruct
This is an experimental self-merge using meta-llama/Meta-Llama-3.1-405B-Instruct and created with mergekit.
This is the direct successor of Meta-Llama-3-120B-Instruct, a self-merge of Llama 3 70B that produced a decent 120B model for tasks like creative writing.
I tweaked the range of duplicated layers to hopefully make a sensible model. Use it at your own risk!
🔍 Applications
I recommend using this model for creative writing with the Llama 3 chat template.
⚡ Quantization
TBD.
🏆 Evaluation
TBD.
🧩 Configuration
This model was merged using the passthrough merge method. The following YAML configuration was used to produce this model:
slices:
- sources:
- layer_range: [0, 42]
model: meta-llama/Meta-Llama-3.1-405B-Instruct
- sources:
- layer_range: [21, 63]
model: meta-llama/Meta-Llama-3.1-405B-Instruct
- sources:
- layer_range: [42, 84]
model: meta-llama/Meta-Llama-3.1-405B-Instruct
- sources:
- layer_range: [63, 105]
model: meta-llama/Meta-Llama-3.1-405B-Instruct
- sources:
- layer_range: [84, 126]
model: meta-llama/Meta-Llama-3.1-405B-Instruct
merge_method: passthrough
dtype: bfloat16
Here is the code I've used to generate the config and calculate the number of layers/parameters after passthrough:
def generate_yaml_config(range_size, total_layers, nb_parameters):
new_size = total_layers + total_layers - range_size
new_param = (nb_parameters / total_layers) * new_size
print(f"New size = {new_size} layers")
print(f"New parameters = {new_param:.2f}B")
yaml_str = "slices:\n"
for i in range(0, round(total_layers - range_size + 1), range_size // 2):
start = i
end = min(start + range_size, total_layers)
yaml_str += f"- sources:\n"
yaml_str += f" - layer_range: [{start}, {end}]\n"
yaml_str += f" model: meta-llama/Meta-Llama-3.1-405B-Instruct\n"
yaml_str += "merge_method: passthrough\n"
yaml_str += "dtype: bfloat16\n"
print(yaml_str)
return new_size, new_param
# Example usage
new_size, new_param = generate_yaml_config(42, 126, 410)
new_size, new_param = generate_yaml_config(105, new_size, new_param)