beberik/Nyxene-v3-11B

Description

This repo contains bf16 files of Nyxene-v1-11B. Just new version with some new things.

Model used

• Intel/neural-chat-7b-v3-3-Slerp
• AIDC-ai-business/Marcoroni-7B-v3
• rwitz/go-bruins-v2
• chargoddard/loyal-piano-m7-cdpo

Prompt template

Just use chatml.

The secret sauce

go-bruins-loyal-piano-11B :

slices:
  - sources:
    - model: rwitz/go-bruins-v2
      layer_range: [0, 24]
  - sources:
    - model: chargoddard/loyal-piano-m7-cdpo
      layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16

neural-marcoroni-11B :

slices:
  - sources:
    - model: AIDC-ai-business/Marcoroni-7B-v3
      layer_range: [0, 24]
  - sources:
    - model: Intel/neural-chat-7b-v3-3-Slerp
      layer_range: [8, 32]

merge_method: passthrough
dtype: bfloat16

Nyxene-11B :

slices:
  - sources:
      - model: "./go-bruins-loyal-piano-11B"
        layer_range: [0, 48]
      - model: "./neural-marcoroni-11B"
        layer_range: [0, 48]
merge_method: slerp
base_model: "./go-bruins-loyal-piano-11B"
parameters:
  t:
    - filter: lm_head 
      value: [0.5]
    - 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: [0.5]
    - 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

Metric	Value
Avg.	70.72
AI2 Reasoning Challenge (25-Shot)	69.62
HellaSwag (10-Shot)	85.33
MMLU (5-Shot)	64.75
TruthfulQA (0-shot)	60.91
Winogrande (5-shot)	80.19
GSM8k (5-shot)	63.53