Meta-Llama-3-8B-InitializedEmbeds

This is just Llama-3-8B with the embeddings for special tokens copied from the Instruct version. Should behave pretty much identically to the base model, but with less glossolalia when it encounters <|start_header_id|> and the like.

I'm using this as a base to fine tune. Having these embeddings reasonable instead of randomly initialized should give a smoother start.

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

Merge Details

Merge Method

This model was merged using the linear merge method.

Models Merged

The following models were included in the merge:

• NousResearch/Meta-Llama-3-8B-Instruct
• NousResearch/Meta-Llama-3-8B

Configuration

The following YAML configuration was used to produce this model:

merge_method: linear
dtype: float32
out_dtype: bfloat16
models:
  - model: NousResearch/Meta-Llama-3-8B
    parameters:
      weight: 1.0
  - model: NousResearch/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.0
tokenizer:
  source: NousResearch/Meta-Llama-3-8B-Instruct
  tokens:
    <|start_header_id|>:
      source: NousResearch/Meta-Llama-3-8B-Instruct
      force: true
    <|end_header_id|>:
      source: NousResearch/Meta-Llama-3-8B-Instruct
      force: true
    <|eot_id|>:
      source: NousResearch/Meta-Llama-3-8B-Instruct
      force: true
    <|end_of_text|>:
      source: NousResearch/Meta-Llama-3-8B-Instruct
      force: true
    <|begin_of_text|>:
      source: NousResearch/Meta-Llama-3-8B-Instruct
      force: true