SUS-Bagel-200K-DARE-Test

Just an experiment to try and extend the context of SUS, a 4K Yi model, and DPO Bagel, which breaks down after 4K context. Yi 4K was used as a base (even for bagel which is technically a Yi 200K model), and Yi 200K is merged in with a density of 1.

I wanted to include Hermes 34B, but something funky about its tokenizer breaks mergekit.

A component of another merge. Auto generated mergekit description below:

---

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

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama as a base.

Models Merged

The following models were included in the merge:

• /home/alpha/Models/Raw/SUSTech_SUS-Chat-34B
• /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
• /home/alpha/Models/Raw/jondurbin_bagel-34b-v0.2
• /home/alpha/Models/Raw/jondurbin_bagel-dpo-34b-v0.2

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
  # No parameters necessary for base model
  - model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
    parameters:
      weight: 0.5
      density: 1
  - model: /home/alpha/Models/Raw/SUSTech_SUS-Chat-34B
    parameters:
      weight: 0.2
      density: 0.12
  - model: /home/alpha/Models/Raw/jondurbin_bagel-dpo-34b-v0.2
    parameters:
      weight: 0.2
      density: 0.15
  - model: /home/alpha/Models/Raw/jondurbin_bagel-34b-v0.2
    parameters:
      weight: 0.1
      density: 0.12
merge_method: dare_ties
tokenizer_source: union
base_model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
parameters:
  int8_mask: true
dtype: bfloat16