Steelskull/Etheria-55b-v0.1

Merge Details

An attempt to make a functional goliath style merge to create a [Etheria] 55b-200k with two yi-34b-200k models.

due to the merge it 'theoretically' should have a context of 200k but I recommend starting at 32k and moveing up, as it is unknown (at this time) what the merge has done to the context length.

This is a merge of both VerA and VerB of Etheria-55b (There numbers were surprisingly good), I then created a sacrificial 55B out of the most performant yi-34b-200k Model and performed a Dare_ties merge and equalize the model into its current state.

recommended settings and Prompt Format:

Ive tested it up to 32k context using exl2 using these settings:

    "temp": 0.7,
    "temperature_last": true,
    "top_p": 1,
    "top_k": 0,
    "top_a": 0,
    "tfs": 1,
    "epsilon_cutoff": 0,
    "eta_cutoff": 0,
    "typical_p": 1,
    "min_p": 0.1,
    "rep_pen": 1.1,
    "rep_pen_range": 8192,
    "no_repeat_ngram_size": 0,
    "penalty_alpha": 0,
    "num_beams": 1,
    "length_penalty": 1,
    "min_length": 0,
    "encoder_rep_pen": 1,
    "freq_pen": 0,
    "presence_pen": 0,
    "do_sample": true,
    "early_stopping": false,
    "add_bos_token": false,
    "truncation_length": 2048,
    "ban_eos_token": true,
    "skip_special_tokens": true,
    "streaming": true,
    "mirostat_mode": 0,
    "mirostat_tau": 5,
    "mirostat_eta": 0.1,

Prompt format that work well

ChatML & Alpaca

Merge Method

This model was merged using the DARE TIES merge method using Merged-Etheria-55b as a base.

Configuration

The following YAML configuration was used to produce this model:

base_model: Merged-Etheria-55b
models:
  - model: Sacr-Etheria-55b
    parameters:
      weight: [0.22, 0.113, 0.113, 0.113, 0.113, 0.113]
      density: 0.61
  - model: Merged-Etheria-55b
    parameters:
      weight: [0.22, 0.113, 0.113, 0.113, 0.113, 0.113]
      density: 0.61
merge_method: dare_ties
tokenizer_source: union
parameters:
  int8_mask: true
dtype: bfloat16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here