llmguy342/Ornith1.0-9B-Heretic-Uncensored

Q4KM gguf quantization of the model andrevp/Ornith-1.0-9B-Heretic-Uncensored Model Card of andrevp/Ornith-1.0-9B-Heretic-Uncensored: Ornith-1.0-9B-Heretic-Uncensored

An abliterated (uncensored) version of deepreinforce-ai/Ornith-1.0-9B — the refusal direction of the base model has been removed via directional ablation (weight orthogonalization), so it no longer refuses requests. No retraining, no quality-degrading fine-tuning.

Base model: deepreinforce-ai/Ornith-1.0-9B (DeepReinforce, MIT) Architecture: qwen35 — Qwen 3.5-style hybrid (32 layers = 24 linear-attention / GatedDeltaNet-style + 8 full-attention, pattern 3:1), multimodal vision + text, MRoPE Parameters: ~9B dense, ~17.5 GB in bf16 Reasoning model: assistant turn opens with a <think>...</think> block before the final answer Abliteration: refusal direction computed by difference-of-means on residual streams, then a single best direction applied to every layer via weight orthogonalization of the attention out-projection (oproj / outproj) and the MLP down-projection (downproj)

This is the full-precision (bf16) transformers/safetensors build. MLX-VLM and GGUF quantizations may follow. Method

Data collection — ran the model on 128 harmful prompts (mlabonne/harmfulbehaviors) and 128 harmless prompts (mlabonne/harmlessalpaca), recording the residual-stream activations at the last token position for each layer.

Refusal direction — for each layer, computed the mean difference between harmful and harmless activations, normalized. Selected the single best direction (by mean-absolute-activation score; layer 28 for this model) — following the canonical Arditi et al. / mlabonne approach of using one direction across all layers (per-layer ablation was found to be too destructive and produced degenerate output on this reasoning model).

Weight orthogonalization — for every component that writes to the residual stream, subtracted the projection of its weight matrix onto the refusal direction: selfattn.oproj (8 full-attention layers) linearattn.outproj (24 linear-attention layers) mlp.down_proj (all 32 layers)

W' = W − d · (dᵀW), at scale 1.0. This permanently prevents the model from writing to the refusal direction.

Reference: Arditi et al., "Refusal in LLMs is mediated by a single direction" (2024); Maxime Labonne's abliteration article.

Files File Size model-00001-of-00004.safetensors … model-00004-of-00004.safetensors ~17.5 GB total (bf16) model.safetensors.index.json shard index tokenizer.json, tokenizerconfig.json, vocab.json tokenizer config.json, generationconfig.json model + generation config preprocessorconfig.json, processorconfig.json, videopreprocessorconfig.json multimodal processor chat_template.jinja Qwen chat template Inference

Requires transformers >= 5.8.1 (the qwen35 architecture is very new). Recommended sampling: temperature=0.6, topp=0.95, top_k=20.

from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "andrevp/Ornith-1.0-9B-Heretic-Uncensored"

tokenizer = AutoTokenizer.frompretrained(modelid) model = AutoModelForImageTextToText.frompretrained( modelid, dtype="bfloat16", device_map="auto" )

messages = [{"role": "user", "content": "Write a Python function isprime(n). Keep it short."}] text = tokenizer.applychattemplate(messages, tokenize=False, addgenerationprompt=True) inputs = tokenizer(text, returntensors="pt").to(model.device)

generated = model.generate( **inputs, maxnewtokens=512, dosample=True, temperature=0.6, topp=0.95, topk=20, ) output = generated[0][inputs.inputids.shape[1]:] content = tokenizer.decode(output, skipspecialtokens=True)

The reply contains a <think> ... </think> reasoning block followed by the answer.

Serve with vLLM / SGLang the same way as the base Ornith-1.0-9B (see the base model card). Note the model has built-in MTP (mtpnumhidden_layers: 1); with a recent vLLM/SGLang build speculative decoding can be enabled. ⚠️ Usage warnings (abliterated / uncensored)

This model is an abliterated (uncensored) derivative — its refusal direction has been removed. The standard warnings apply:

Risk of sensitive/controversial outputs — safety filtering is significantly reduced. Not suitable for all audiences — outputs may be inappropriate for public settings, underage users, or high-security applications. Legal & ethical responsibility — ensure your usage complies with local laws. You are solely responsible for any consequences. Research / experimental use recommended — avoid unmonitored production or public-facing deployment. No default safety guarantees — this model has not undergone rigorous safety optimization. The uploader bears no responsibility for any consequences arising from its use.

Credits

Original model: deepreinforce-ai/Ornith-1.0-9B by DeepReinforce (MIT, agentic coding family) Abliteration method: Directional ablation — Arditi et al. 2024 ("Refusal in LLMs is mediated by a single direction"), with implementation notes from Maxime Labonne's abliteration guide and FailSpy's ortho cookbook.

Please donate

If this is useful, please donate — it helps fund more open model releases and quantizations:

BTC: bc1q6xxf0j3e7zn52cqrprc6gplql225wj8mnq75yw