beberik/Nyxene-v1-11B

Description

This repo contains bf16 files of Nyxene-v1-11B. Same as the previous version but I used newer models and tried to repeat what I experimented with when there were older models.

Model used

• berkeley-nest/Starling-LM-7B-alpha
• openaccess-ai-collective/DPOpenHermes-7B
• fblgit/juanako-7b-UNA
• chargoddard/loyal-piano-m7
• argilla/notus-7b-v1

I added a new model because after the same action but using zephyr and dolphin the model turned out to be more creative.

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 with 1% of notus :

slices:
  - sources:
      - model: chargoddard/loyal-piano-m7
        layer_range: [0, 48]
      - model: argilla/notus-7b-v1
        layer_range: [0, 48]
merge_method: slerp
base_model: argilla/notus-7b-v1
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: 0.99 # fallback for rest of tensors
dtype: bfloat16

loyal-piano-juanako-11B :

slices:
  - sources:
    - model: fblgit/juanako-7b-UNA
      layer_range: [0, 24]
  - sources:
    - model: chargoddard/loyal-piano-m7
      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-juanako-11B
        layer_range: [0, 48]
      - model: Starling-NeuralHermes-11B
        layer_range: [0, 48]
merge_method: slerp
base_model: dolphin-juanako-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.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	67.58
AI2 Reasoning Challenge (25-Shot)	67.49
HellaSwag (10-Shot)	84.52
MMLU (5-Shot)	65.12
TruthfulQA (0-shot)	57.28
Winogrande (5-shot)	79.01
GSM8k (5-shot)	52.08