Llama-3-13B-Instruct

This is a merge of pre-trained language models created using mergekit.

The goal was to create a Llama 3 13B model to have a "mid" sized model which Meta has released in the past, but I would consider this a base model to be further finetuned on. Surprisingly, it is usable for chat and storywriting with Llama 3 Instruct template, though it does occasionally have some grammatical quirks like L3-120B.

Logical ability (programming, math, science, etc.) has been deteriorated by the merge process.

Use no repetition penalty or <1.05 or it might go a bit haywire, other than that, it is suitable for writing use. I have not tested it against L3 8B in that regard.

Finetuned Version

A finetuned version of this model can be found at elinas/Llama-3-13B-Instruct-ft which seems to improve performance.

Merge Details

Merge Method

This model was merged using the passthrough merge method.

Models Merged

The following models were included in the merge:

• meta-llama/Meta-Llama-3-8B-Instruct

Configuration

The following YAML configuration was used to produce this model:

dtype: bfloat16
merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 10]
    model: meta-llama/Meta-Llama-3-8B-Instruct
- sources:
  - layer_range: [5, 15]
    model: meta-llama/Meta-Llama-3-8B-Instruct
- sources:
  - layer_range: [10, 20]
    model: meta-llama/Meta-Llama-3-8B-Instruct
- sources:
  - layer_range: [15, 25]
    model: meta-llama/Meta-Llama-3-8B-Instruct
- sources:
  - layer_range: [20, 25]
    model: meta-llama/Meta-Llama-3-8B-Instruct
- sources:
  - layer_range: [22, 32]
    model: meta-llama/Meta-Llama-3-8B-Instruct

Model Evaluation

TBD - submitted