A LoRA was applied to "abliterate" refusals in meta-llama/Meta-Llama-3.1-8B-Instruct. The result appears to work despite the LoRA having been derived from Llama 3 instead of Llama 3.1, which implies that there is significant feature commonality between the 3 and 3.1 models.

The LoRA was extracted from failspy/Meta-Llama-3-8B-Instruct-abliterated-v3 using meta-llama/Meta-Llama-3-8B-Instruct as a base.

Built with Llama.

Merge Details

Merge Method

This model was merged using the task arithmetic merge method using meta-llama/Meta-Llama-3.1-8B-Instruct + grimjim/Llama-3-Instruct-abliteration-LoRA-8B as a base.

Configuration

The following YAML configuration was used to produce this model:

yaml

base_model: meta-llama/Meta-Llama-3.1-8B-Instruct+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 32]
    model: meta-llama/Meta-Llama-3.1-8B-Instruct+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
    parameters:
      weight: 1.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	22.95
IFEval (0-Shot)	48.70
BBH (3-Shot)	29.42
MATH Lvl 5 (4-Shot)	12.39
GPQA (0-shot)	8.50
MuSR (0-shot)	9.26
MMLU-PRO (5-shot)	29.46

grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter

Llama-3.1-8B-Instruct-abliterated_via_adapter

Merge Details

Merge Method

Configuration

Open LLM Leaderboard Evaluation Results

grimjim/
Llama-3.1-8B-Instruct-abliterated_via_adapter