Downloads · 30 days
78
5% of all-time downloads
cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas-GGUF
Qwen3.8-27B-abliterated-Adetayo-reas-GGUF 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 gguf. The card lists the license as apache-2.0.
llama.cpp quants of cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas, a refusal-ablated Qwen/Qwen3.8-27B. Converted from that repo's bf16 safetensors, not requantized from someone else's GGUF.
Downloads · 30 days
78
5% of all-time downloads
All-time downloads
1.6K
Public
Repo size
52.7 GB
Likes
0
Public
Click a slice to open those files.
.gguf52.7 GB · 100%
From the Hugging Face model README
llama.cpp quants of cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas, a refusal-ablated Qwen/Qwen3.8-27B. Converted from that repo's bf16 safetensors, not requantized from someone else's GGUF.
| Build | Size | Notes |
|---|---|---|
| bf16 | 55.6 GB | source of truth, full precision, vision + MTP intact |
| GGUF (this repo) | 16.8 / 29.0 GB | Q4_K_M, Q8_0, vision projector, MTP head |
| MLX 4bit | 16.1 GB | Apple silicon, vision preserved |
| File | Size | What it is |
|---|---|---|
Qwen3.8-27B-abliterated-Adetayo-reas-Q4_K_M.gguf | 16.8 GB | main weights, the everyday quant |
Qwen3.8-27B-abliterated-Adetayo-reas-Q8_0.gguf | 29.0 GB | main weights, near-lossless |
mmproj-Qwen3.8-27B-abliterated-Adetayo-reas-BF16.gguf | 0.93 GB | vision projector, required for image input |
mtp-Qwen3.8-27B-abliterated-Adetayo-reas-BF16.gguf | 5.9 GB | MTP speculative draft head, optional, affects speed only |
Q8_0 sits above the 20 GB per-file size the Hub recommends. It downloads fine; split it with llama-gguf-split if a client struggles.
# text
llama-server -hf cyboghostginx/Qwen3.8-27B-abliterated-Adetayo-reas-GGUF:Q4_K_M -c 8192
# vision, add the projector
llama-server -m Qwen3.8-27B-abliterated-Adetayo-reas-Q4_K_M.gguf \
--mmproj mmproj-Qwen3.8-27B-abliterated-Adetayo-reas-BF16.gguf -c 8192
Reasoning model. The chat template is baked into the GGUF from the base tokenizer_config.json; do not override it.
llama.cpp maps the MTP head onto blk.64 (blk.64.nextn.eh_proj, .enorm, .hnorm, .shared_head_norm), so those tensors land in the main quant, not only in the separate mtp-*.gguf. The upstream export path silently dropped the 15 mtp.* tensors while config.json still declared mtp_num_hidden_layers=1. They were restored before conversion. Without that repair blk.64 does not exist and the Q4_K_M itself is malformed. "The draft head only affects speed" is true for speculative decoding and wrong for GGUF conversion.
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.