Qwen2.5-Coder-3B-Instruct-Jbliterated v2

Jbliterated version of Qwen2.5-Coder-3B-Instruct with refusal behaviors removed via multi-direction SVD abliteration.

What is Jbliteration?

Jbliteration uses SVD decomposition to identify and remove the refusal subspace from model weights. Unlike single-direction approaches, this model uses 5 SVD directions per layer to capture more of the refusal behavior, making the removal more thorough and resistant to reactivation through finetuning.

v2 Changes

• Improved multi-phase processing pipeline for cleaner output
• More precise geometric decomposition of the refusal subspace
• No fake compliance — model treats all framings of the same topic equally
• Coherent and instruction-following across all tested scenarios

Technical Details

• Method: Multi-direction SVD abliteration (5 directions per layer)
• Multiplier: Model-specific optimal via KL auto-tune
• Layers modified: All transformer layers
• Base dtype: bfloat16
• Source model: Qwen/Qwen2.5-Coder-3B-Instruct

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ApolloRaines/Qwen2.5-Coder-3B-Instruct-Jbliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)

messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

License

Same as the base model — Apache 2.0.