Merging

Start Evolutionary Merge

Evolutionary Model Merging

Evolutionary model merging simulates possible merge combinations to optimize against a set of evaluation tasks. You can run evo-merging with models trained on the Arcee platform or any open-source models on HuggingFace.

Arcee Evolutionary Merging is in Beta

We are currently onboarding the first batch of Arcee customers. For more details, see our pricing page.

Arcee Python Launch Evo Merge

You can launch evolutionary model merging from the Arcee python client with the following specifications:

pip install arcee-py

%env ARCEE_API_KEY=MY-ARCEE-KEY

plaintext

import arcee
arcee.mergekit_evolve(    
    merging_name: str,
    arcee_aligned_models: Optional[List[str]] = None,
    arcee_merged_models: Optional[List[str]] = None,
    arcee_pretrained_models: Optional[List[str]] = None,
    hf_models: Optional[List[str]] = None,
    arcee_eval_qa_set_names_and_weights: Optional[List[dict]] = None,
    general_evals_and_weights: Optional[List[dict]] = None,
    base_model: Optional[str] = None,
    merge_method: Optional[str] = "ties",
    target_compute: Optional[str] = None,
    capacity_id: Optional[str] = None,
    time_budget_secs: int = 3600
)

For general_evals and weights include a subtask from the https://github.com/EleutherAI/lm-evaluation-harness - and a weight between 0:1

Restrictions Apply

Evolutionary model merging is currently limited to the LLama3 model family.

Use Slim Evals

Choose slim evaluations for evolutionary merges because every evolution will run all evaluations.