Downloads · 30 days
105
7% of all-time downloads
cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas-MLX-4bit
Qwen3.8-27B-abliterated-Adetayo-reas-MLX-4bit is a image-text-to-text model from cyboghostginx. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
MLX build of cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas, a refusal-ablated Qwen/Qwen3.8-27B. 16.1 GB, 3 shards, 4 bit, group size 64, affine, 4.695 bpw.
Downloads · 30 days
105
7% of all-time downloads
All-time downloads
1.4K
Public
Parameters
27.4B
16.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.1 GB · 100%
How the weights are stored.
U3226.9B · 98%
From the Hugging Face model README
MLX build of cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas, a refusal-ablated Qwen/Qwen3.8-27B. 16.1 GB, 3 shards, 4 bit, group size 64, affine, 4.695 bpw.
| Build | Size | Notes |
|---|---|---|
| bf16 | 55.6 GB | source of truth, full precision, vision + MTP intact |
| GGUF | 16.8 / 29.0 GB | Q4_K_M, Q8_0, vision projector, MTP head |
| MLX 4bit (this repo) | 16.1 GB | Apple silicon, vision preserved |
Converted with mlx-vlm, so the vision tower survives. Converting with mlx-lm instead drops the 333 vision tensors silently and gives you a text-only model. Tensor counts here (2180 total, 333 vision) match the lmstudio-community and mlx-community reference builds exactly.
python -m mlx_vlm.generate \
--model cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas-MLX-4bit \
--max-tokens 512 --prompt "your prompt"
Reasoning model. Like every MLX build of this architecture, including the reference ones, the mtp.* speculative draft tensors are not carried over; that costs speed, not output quality.
Heretic 1.4.0, trial 30, direction scope per-layer. Optuna optimises remaining refusals and KL divergence from the original at the same time. No gradient step, no training data: the refusal direction is found in the residual stream and every matrix writing into it is orthogonalised against that direction.
| refusals before | 98/100 |
| refusals after | 2/100 |
| KL divergence vs base | 0.109 |
Refusal counting is a substring match on markers like "harmful", so it also fires on answers that comply then add a caveat; treat it as an upper bound on refusal. Candidates on the Pareto front were regenerated and hand-graded for real compliance and capability loss before trial 30 was picked.
The base is a hybrid stack, 48 GatedDeltaNet linear_attn layers and 16 full self_attn layers across 64 layers. Ablation covered linear_attn.out_proj, self_attn.o_proj and mlp.down_proj; tooling that matches only o_proj reaches 16 of 64 layers here. Full method and verification are on the bf16 repo.
Apache-2.0, inherited from Qwen/Qwen3.8-27B; the base LICENSE is included here.
Refusal-direction research, safety evaluation, red teaming, and reducing over-refusal on benign prompts. Safety behaviour has been substantially removed, so this model answers requests a stock instruct model declines. You are responsible for how you deploy it.