noctrex/Mistral-7B-Instruct-v0.3-abliterated-GGUF

This is a abliterated version of Mistral-7B-Instruct-v0.3, made using Heretic

The quantizations were created using an imatrix merged from text_en_large and harmful.txt to leverage the abliterated nature of the model.

Performance

Metric	This model	Original model
KL divergence	0.15	0 (by definition)
Refusals	3/100	85/100

Analysis against the original model:

• Total Tensors: 291
• Tensors with Diffs: 47 (16.2%)
• Average % Diff: 6.99%
• Median % Diff: 0.00%
• Min/Max % Diff: 0.00% / 46.77%
• Std Dev % Diff: 15.97%
• Skewness % Diff: 1.86
• Avg L2 Norm: 144619.57
• Tensors with >5% diff: 47
• Top differences:
  • blk.14.attn_output.weight ((4096, 8192), L2: 669167.94): 46.77%
  • blk.13.attn_output.weight ((4096, 8192), L2: 667456.52): 46.51%
  • blk.16.attn_output.weight ((4096, 8192), L2: 667644.60): 46.46%
  • blk.12.attn_output.weight ((4096, 8192), L2: 664339.15): 46.03%
  • blk.15.attn_output.weight ((4096, 8192), L2: 664117.46): 45.94%