SteelStorage/Aura-Llama-Abliterated

Aura-llama-3-Abliterated

Now that the cute anime girl has your attention.

UPDATE: Model is now using the abliterated version of meta llama 3 8b

Aura-llama is using the methodology presented by SOLAR for scaling LLMs called depth up-scaling (DUS), which encompasses architectural modifications with continued pretraining. Using the solar paper as a base, I integrated Llama-3 weights into the upscaled layers, and In the future plan to continue training the model.

Aura-llama is a merge of the following models to create a base model to work from:

• failspy/Llama-3-8B-Instruct-abliterated
• failspy/Llama-3-8B-Instruct-abliterated

Abliterated Merged Evals (Has Not Been Finetuned):

Aura-llama-Abliterated

• Avg: ?
• ARC: ?
• HellaSwag: ?
• MMLU: ?
• T-QA: ?
• Winogrande: ?
• GSM8K: ?

Non Abliterated Merged Evals (Has Not Been Finetuned):

Aura-llama-Original

• Avg: 63.13
• ARC: 58.02
• HellaSwag: 77.82
• MMLU: 65.61
• T-QA: 51.94
• Winogrande: 73.40
• GSM8K: 52.01

🧩 Configuration

dtype: bfloat16
merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 12]
    model: failspy/Llama-3-8B-Instruct-abliterated
- sources:
  - layer_range: [8, 20]
    model: failspy/Llama-3-8B-Instruct-abliterated
- sources:
  - layer_range: [16, 28]
    model: failspy/Llama-3-8B-Instruct-abliterated
- sources:
  - layer_range: [24, 32]
    model: failspy/Llama-3-8B-Instruct-abliterated
        

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	53.46
AI2 Reasoning Challenge (25-Shot)	49.23
HellaSwag (10-Shot)	72.27
MMLU (5-Shot)	55.71
TruthfulQA (0-shot)	46.63
Winogrande (5-shot)	69.30
GSM8k (5-shot)	27.60