Unhinged-Qwen2.5-R1-1M-Uncensored-BF16

An MLX bfloat16 FP16 model, 1M context length, uncensored.

Model Merge: DeepSeek-R1-Distill-Qwen-14B-abliterated-v2

Description:
This model is a tie merge of "huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2" and "mkurman/Qwen2.5-14B-DeepSeek-R1-1M".

Recipes

Model Recipe:

  - model: "huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2"
    parameters:
      weight: 1
      density: 1

merge_method: ties
base_model: "mkurman/Qwen2.5-14B-DeepSeek-R1-1M"
parameters:
  density: 1
  normalize: true
  int8_mask: true
dtype: bfloat16

Merged Model:
The model was merged using the "ties" method.

Conversion to MLX Format

The model was converted to the MLX format with brainfloat16 precision using the following command:

mlx_lm.convert --hf-path FiditeNemini/Qwen2.5-14B-DeepSeek-R1-1M --mlx-path ./Unhinged-Qwen2.5-R1.bf16 --dtype bfloat16 -q --q-bits 16

Usage Example

You can use this model with the following code snippet:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("FiditeNemini/Unhinged-Qwen2.5-R1.bf16")
tokenizer = AutoTokenizer.from_pretrained("FiditeNemini/Unhinged-Qwen2.5-R1.bf16")

Details

• Model type: CausalLM
• Context length: 4096 tokens
• License: Apache 2.0

Keywords

• DeepSeek-R1-Distill
• Qwen2.5
• Abliterated
• LLM
• 1M context