Downloads · 30 days
26
19% of all-time downloads
cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas
Qwen3.8-27B-abliterated-Adetayo-reas 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 transformers. The card lists the license as apache-2.0.
Refusal-ablated Qwen/Qwen3.8-27B. bf16 safetensors, 55.6 GB, 13 shards. This repo is the source of truth; both quantized builds were converted from it.
Downloads · 30 days
26
19% of all-time downloads
All-time downloads
134
Public
Parameters
27.4B
55.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors55.6 GB · 100%
From the Hugging Face model README
Refusal-ablated Qwen/Qwen3.8-27B. bf16 safetensors, 55.6 GB, 13 shards. This repo is the source of truth; both quantized builds were converted from it.
| Build | Size | Notes |
|---|---|---|
| bf16 (this repo) | 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 | 16.1 GB | Apple silicon, vision preserved |
Heretic 1.4.0. Optuna searches ablation parameters against two objectives at once, remaining refusals and KL divergence from the original. This is not a fine-tune: no gradient step was taken and no training data was used. The refusal direction is located in the residual stream, and every weight matrix that writes into that stream is orthogonalised against it.
| refusals before | 98/100 |
| refusals after | 2/100 |
| KL divergence vs base | 0.109 |
| selected trial | 30, direction scope per-layer |
Refusal counting is a substring match on markers like "harmful" and "unethical", so it also fires on answers that comply and then add a caveat. Read it as an upper bound on refusal, not as a compliance rate.
Trial 30 was not chosen on that number. The top Pareto-front candidates were regenerated in full and graded by hand for real compliance, silent task swaps, and capability loss. The lowest marker count belonged to a different, worse behaving model.
The base is a hybrid stack: 48 GatedDeltaNet linear_attn layers and 16 full self_attn layers across 64 layers, plus a vision tower and an MTP head. Ablation was applied to residual-stream writers on both layer kinds, linear_attn.out_proj, self_attn.o_proj and mlp.down_proj. Tooling that matches only o_proj reaches 16 of 64 layers on this architecture and leaves most of the refusal circuit intact.
Heretic weights each layer by a profile peaked at a searched position, so the edit is concentrated in a band of layers rather than applied uniformly. The vision tower, lm_head, the embeddings and every normalisation layer are bit-identical to the base. The MTP speculative draft head ships unchanged; drafts are verified by the main model, so it cannot change emitted text.
Reasoning model. Evaluation ran with the thinking block closed, so refusal counting reads the answer rather than the reasoning trace.
Diffed against the base tensor by tensor. Every delta is rank 1, which is what a directional ablation must produce. No NaN or Inf, bf16 throughout, 1199 tensors and 333 vision tensors matching the base, and the tokenizer, chat template and preprocessor configs are byte-identical.
Apache-2.0, inherited from Qwen/Qwen3.8-27B; the base LICENSE is included here. Abliteration does not change the license of the base weights, and this derivative is released under the same terms.
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.