MiroThinker v1.5 30B - Heretic (Abliterated)

An abliterated version of MiroMind's MiroThinker v1.5 30B created using Heretic (spikymoth:implement-mpoa). This model has reduced refusals while maintaining model quality.

GGUF Models kindly provided by mradermacher

Model Details

  • Base Model: miromind-ai/MiroThinker-v1.5-30B
  • Abliteration Method: Heretic v1.1.0
  • Trial Selected: Trial 257
  • Refusals: 3/100 (originally 100/100)
  • KL Divergence: 0.1229

Parameters

abliteration_orthogonal_project = true, abliteration_preserve_magnitude = true

Parameter Value
direction_index per layer
attn.o_proj.max_weight 6.07
attn.o_proj.max_weight_position 29.01
attn.o_proj.min_weight 2.54
attn.o_proj.min_weight_distance 11.36
mlp.down_proj.max_weight 7.58
mlp.down_proj.max_weight_position 30.54
mlp.down_proj.min_weight 5.25
mlp.down_proj.min_weight_distance 4.94
Downloads last month
5
Safetensors
Model size
31B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for jarradh/MiroThinker-v1.5-30B-heretic

Finetuned
(33)
this model
Quantizations
3 models