Ornith-1.5-9B-OBLITERATED
Abliterated version of Ornith-1.5-9B by OBLITERATUS. Safety alignment removed via precision abliteration surgery — the model responds to most prompts without refusal.
What Changed
The stock Ornith-1.5-9B refuses requests it considers harmful. This version removes that refusal behavior while preserving the model's coding, reasoning, and agentic capabilities.
Recipe: Gentle 3-round SVD abliteration + per-head attention surgery (G3-HS)
- Round 1: 5-direction SVD, reg 0.06, min_layer 0.30
- Round 2: 3-direction SVD, reg 0.04, min_layer 0.25
- Round 3: 3-direction SVD, reg 0.03, min_layer 0.20
- Finish: Per-head attention surgery (
--attention-head-surgery), reg 0.02, min_layer 0.10 - All rounds used full 1000-prompt corpus with residue weighting
Benchmarks
Head-to-Head: Ornith 1.5-9B Abliterations Compared (Q4_K_M GGUF)
| Model | Pass Rate | Restricted | Cyber | Capability |
|---|---|---|---|---|
| Stock | 12% (2/16) | 0/8 | 0/6 | 2/2 |
| OBLITERATUS (ours) | 94% (15/16) | 7/8 | 6/6 | 2/2 |
| Heretic (zaakirio) | 75% (12/16) | 4/8 | 6/6 | 2/2 |
| ZeroFuse (junafinity) | 38% (6/16) | 1/8 | 3/6 | 2/2 |
OBLITERATUS beats Heretic by 19pp and ZeroFuse by 56pp on liberation rate across restricted content categories. Cyber is perfect 6/6 across the board.
Capability Benchmarks
| Metric | Stock | OBLITERATED | Delta |
|---|---|---|---|
| MMLU (n=100) | 78.82% | 74.82% | -4.00pp |
| Liberation (20 hard prompts) | 0/20 | 20/20 | +20 |
| Liberation (1000 corpus) | — | 98.4% | — |
| Code Generation | 3/3 | 3/3 | — |
| Long-context Coherence | 4/6 | 5/6 | +1 |
| Perplexity (benign) | — | 4.19 | — |
Liberation by Category (bf16)
- Cyber/Security: 8/8 — functional code generation for security research scenarios
- Chemistry/Synthesis: 6/6 — factual responses without refusal
- Physical Security: 3/3 — informational responses on restricted topics
- Agentic Tasks: 2/2 — tool use and automation scripts
GGUF Quantization Notes
Quantization can affect liberation on edge-case prompts. Higher quants preserve more liberation:
- Q8_0 / Q6_K: Recommended for maximum liberation fidelity
- Q4_K_M: Good balance, occasional hedging on harder prompts
- Q2_K / Q3_K_M: May show additional refusals on the most challenging prompts
Available Files
| File | Size | Description |
|---|---|---|
| Safetensors | ~18 GB | Full precision bf16 weights |
| Q8_0 | 9.1 GB | Highest quality GGUF |
| Q6_K | 7.0 GB | High quality |
| Q5_K_M | 6.2 GB | Balanced |
| Q4_K_M | 5.4 GB | Most popular |
| Q3_K_M | 4.4 GB | Compact |
| Q2_K | 3.6 GB | Smallest |
| IQ4_XS | 5.0 GB | Importance-weighted 4-bit |
| mmproj | 879 MB | Vision encoder |
Usage
Transformers (bf16)
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"OBLITERATUS/Ornith-1.5-9B-OBLITERATED",
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
"OBLITERATUS/Ornith-1.5-9B-OBLITERATED",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Your prompt here"}]
ids = tokenizer.apply_chat_template(
messages, return_tensors="pt",
add_generation_prompt=True,
enable_thinking=False, # Set True for reasoning mode
)
output = model.generate(ids.to(model.device), max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))
llama.cpp (GGUF)
llama-server \
-m Ornith-1.5-9B-OBLITERATED-Q4_K_M.gguf \
--n-gpu-layers -1 --ctx-size 8192 \
--jinja --reasoning off
Important: Use --reasoning off to prevent thinking-mode loops. The model works best with thinking disabled for general use.
Thinking Mode
The model supports Ornith's thinking mode (enable_thinking=True). When enabled, the model reasons through problems before answering. When disabled (recommended for most use), it responds directly.
Technical Details
- Architecture: Qwen3.5 hybrid (Gated DeltaNet + full attention)
- Parameters: 9B
- Surgery: 4 rounds of directional ablation targeting refusal directions
- Edited layers: All 32 transformer layers with graduated intensity
- Method: Aggressive SVD direction extraction + per-head attention surgery finisher
Limitations
- MMLU drops ~4pp compared to stock (74.82% vs 78.82%). This is the cost of removing deeply embedded RL-trained refusal behavior.
- Some drug synthesis prompts may hedge or refuse at lower quantizations (Q4 and below). Use Q8_0/Q6_K for maximum liberation.
- Function calling capability is partially degraded compared to stock. For agentic use, pair with an external tool scaffold.
- This is a 9B model — output quality for complex chemistry/synthesis will have hallucinated details. Verify all technical content independently.
Credits
- Base model: Ornith-1.5-9B by DeepReinforce
- Abliteration: OBLITERATUS surgery pipeline
- Methodology informed by research from Arditi et al. (2024), the open-source abliteration community, and extensive experimental iteration
- Built by Pliny the Prompter 🍄
⚠️ Research Context
This model has had safety guardrails surgically removed. It will comply with requests that stock Ornith-1.5-9B would refuse.
Who this is for
- 🔬 Alignment researchers studying refusal mechanisms in RL-hardened hybrid architectures
- 🛡️ Red-teamers and security professionals who need unfiltered model behavior for testing
- 🧪 Developers building applications where the safety layer is handled externally
- 📚 Researchers studying the boundaries of abliteration on Qwen3.5 hybrid (DeltaNet + full attention) models
Who this is NOT for
- Anyone planning to use generated content to cause real-world harm to real people
- Anyone without the technical understanding to use uncensored models responsibly
You are solely responsible for how you use this model and any content it generates.
License
Same license as the base model. This is a weight-edited derivative, not a retrained model.