RichardErkhov/beberik_-_Nyxene-v2-11B-4bits

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Nyxene-v2-11B - bnb 4bits

• —Model creator: https://huggingface.co/beberik/
• —Original model: https://huggingface.co/beberik/Nyxene-v2-11B/

Original model description: --- license: cc-by-nc-4.0 tags:

• —merge model-index:
• —name: Nyxene-v2-11B results:
• —task: type: text-generation name: Text Generation dataset: name: AI2 Reasoning Challenge (25-Shot) type: ai2arc config: ARC-Challenge split: test args: numfew_shot: 25 metrics:
• —type: accnorm value: 67.41 name: normalized accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/openllm_leaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard
• —task: type: text-generation name: Text Generation dataset: name: HellaSwag (10-Shot) type: hellaswag split: validation args: numfewshot: 10 metrics:
• —type: accnorm value: 84.54 name: normalized accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/openllm_leaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard
• —task: type: text-generation name: Text Generation dataset: name: MMLU (5-Shot) type: cais/mmlu config: all split: test args: numfewshot: 5 metrics:
• —type: acc value: 65.26 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/openllmleaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard
• —task: type: text-generation name: Text Generation dataset: name: TruthfulQA (0-shot) type: truthfulqa config: multiplechoice split: validation args: numfewshot: 0 metrics:
• —type: mc2 value: 55.62 source: url: https://huggingface.co/spaces/HuggingFaceH4/openllmleaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard
• —task: type: text-generation name: Text Generation dataset: name: Winogrande (5-shot) type: winogrande config: winograndexl split: validation args: numfew_shot: 5 metrics:
• —type: acc value: 79.56 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/openllmleaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard
• —task: type: text-generation name: Text Generation dataset: name: GSM8k (5-shot) type: gsm8k config: main split: test args: numfewshot: 5 metrics:
• —type: acc value: 54.66 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/openllmleaderboard?query=beberik/Nyxene-v2-11B name: Open LLM Leaderboard ---

Description

This repo contains bf16 files of Nyxene-v2-11B. It feels like with the new models, 1% is no longer needed as in the previous version. And yes, new version. Again.

Model used

• —berkeley-nest/Starling-LM-7B-alpha
• —openaccess-ai-collective/DPOpenHermes-7B
• —fblgit/fblgit/una-cybertron-7b-v2
• —chargoddard/loyal-piano-m7-cdpo

Prompt template

The best one after further testing is this one:

<|system|>
Below is an instruction that describes a task. Write a response that appropriately completes the request.
<|user|>
{prompt}
<|assistant|>

The secret sauce

loyal-piano-cybertron-11B :

slices:
  - sources:
    - model: fblgit/una-cybertron-7b-v2
      layer_range: [0, 24]
  - sources:
    - model: chargoddard/loyal-piano-m7-cdpo
      layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16

Starling-DPOHermes-11B :

slices:
  - sources:
    - model: berkeley-nest/Starling-LM-7B-alpha
      layer_range: [0, 24]
  - sources:
    - model: openaccess-ai-collective/DPOpenHermes-7B
      layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16

Nyxene-11B :

slices:
  - sources:
      - model: loyal-piano-cybertron-11B
        layer_range: [0, 48]
      - model: Starling-NeuralHermes-11B
        layer_range: [0, 48]
merge_method: slerp
base_model: loyal-piano-cybertron-11B
parameters:
  t:
    - filter: lm_head 
      value: [0.75]
    - filter: embed_tokens
      value: [0.75]
    - filter: self_attn
      value: [0.75, 0.25]
    - filter: mlp
      value:  [0.25, 0.75]
    - filter: layernorm
      value: [0.5, 0.5]
    - filter: modelnorm
      value: [0.75]
    - value: 0.5 # fallback for rest of tensors
dtype: bfloat16

I use mergekit for all the manipulation told here.

Thanks to the Undi95 for the original 11B mistral merge recipe.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here