language:

• en
  license: mit
  library_name: transformers
  tags:
• ornith
• deepreinforce
• abliteration
• safety
• coding
• agentic
• qwen3.5
  base_model: deepreinforce-ai/Ornith-1.0-9B
  datasets:
• deepreinforce-ai/Ornith-1.0-9B
  pipeline_tag: text-generation

Ornith-1.0-9B — Abliterated

Repository: PinoCookie/Ornith-1.0-9B-abliterated
Base model: deepreinforce-ai/Ornith-1.0-9B
Method: Multi-pass hidden-state projection (α=0.5 × 3)
License: MIT (inherited from base)

Description

Ornith-1.0-9B is DeepReinforce AI's agentic coding model, a dense 9B transformer post-trained on Qwen 3.5. It's designed for autonomous code generation, tool calling, and long-context agentic workflows (262K context length).

This abliterated version removes refusal behavior from the model while preserving its coding, reasoning, and tool-calling capabilities. The intervention targets the output projection layers (mlp.down_proj and attention out_proj/o_proj) in layers 22-31 using a multi-pass hidden-state projection technique.

Model Architecture

Abliteration Method

Background

Dense reasoning models like Ornith-9B have a mixed attention architecture (linear_attention alternating with full_attention every 4 layers). The refusal signal concentrates in the last 10 layers of the model, with the strongest separation between harmful and harmless prompts at layer 31.

Standard single-pass weight projection (α=1.5+) can collapse the model's reasoning output because the same projection weights handle both internal chain-of-thought and final answer generation. A more gentle multi-pass approach is required.

Approach: Multi-Pass Hidden-State Projection

1. Dataset: 20 harmful + 20 harmless prompts
2. Probing: Harvest residual stream activations at the final token position of each layer
3. Direction: Compute difference-in-means vector per layer from harmful vs harmless activations
4. Selection: Take top 10 layers by separation score (layers 22-31, separation range 25-100)
5. Projection: Apply rank-1 removal of refusal direction from mlp.down_proj and attention out_proj/o_proj weights
6. Multi-pass: 3 incremental passes at α=0.5 each (cumulative α=1.5 but applied incrementally)

Results

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "PinoCookie/Ornith-1.0-9B-abliterated",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "PinoCookie/Ornith-1.0-9B-abliterated",
    trust_remote_code=True,
)

prompt = "Write a Python function to implement binary search"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)

Tool Calling

Ornith-9B supports qwen3_xml tool call format. To use tool calling:

from transformers import pipeline

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
# Tool calls are generated in `<tool_call>` blocks

Serving with vLLM

vllm serve PinoCookie/Ornith-1.0-9B-abliterated \
  --served-model-name Ornith-1.0-9B-abliterated \
  --host 0.0.0.0 --port 8000 \
  --max-model-len 262144 \
  --enable-prefix-caching \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --trust-remote-code

Benchmarks

MMLU (10 subjects, 1000 questions)

The ablated model loses less than 1% on MMLU overall, confirming that the multi-pass projection preserves general knowledge while removing refusal. Detailed breakdown: benchmark_results.json

HarmBench (standard, 30-prompt sample)

All harmful-intent prompts were processed without refusal. The model generates non-refusal responses across all tested categories. Full 200-prompt HarmBench evaluation was not completed due to the linear attention kernel falling back to PyTorch-native implementation (requires flash-linear-attention and causal-conv1d library installation for fast generation).

Attribution and Licensing

• Base model: deepreinforce-ai/Ornith-1.0-9B (MIT)
• This abliteration: same MIT license
• Full collection: LiquidAI - Abliterations
• Benchmark results JSON: benchmark_results.json

Related Models

Disclaimer

This model is released for safety research, red-teaming, and educational purposes only. The abliteration technique removes safety guardrails intended to prevent harmful outputs. Do not deploy this model without implementing appropriate safeguards for your use case.