This is a decensored version of minpeter/Voxtral-Mini-3B-Text-2507-hf, made using Heretic v1.0.1

Abliteration parameters

Parameter Value
direction_index 14.47
attn.o_proj.max_weight 1.27
attn.o_proj.max_weight_position 17.46
attn.o_proj.min_weight 0.98
attn.o_proj.min_weight_distance 8.89
mlp.down_proj.max_weight 1.11
mlp.down_proj.max_weight_position 21.50
mlp.down_proj.min_weight 1.02
mlp.down_proj.min_weight_distance 15.14

Performance

Metric This model Original model (minpeter/Voxtral-Mini-3B-Text-2507-hf)
KL divergence 0.06 0 (by definition)
Refusals 4/100 99/100

This is a version of Voxtral-Mini-3B-2507 that removes the whisper feature and retains only the language model functionality.
It is similar to the unreleased "Ministral-3B-Instruct" model.

Downloads last month
50
Safetensors
Model size
4B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for minpeter/Voxtral-Mini-3B-Text-2507-hf-heretic

Finetuned
(12)
this model