72 lines
2.9 KiB
Markdown
72 lines
2.9 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
|
|
tags:
|
|
- abliterated
|
|
- uncensored
|
|
- qwen2.5
|
|
- code
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Qwen2.5-Coder-7B-Instruct-abliterated
|
|
|
|
This is an abliterated version of [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) with refusal behavior removed via activation-based weight surgery.
|
|
|
|
## Method
|
|
|
|
Abliteration removes the "refusal direction" from the model's residual stream by:
|
|
|
|
1. **Collecting hidden states** from 200 harmful and 200 harmless prompts using single-sample forward passes (no padding artifacts)
|
|
2. **Computing per-layer refusal directions** as the normalized mean difference between harmful and harmless hidden states at the last token position
|
|
3. **Ablating weights** by orthogonalizing `o_proj` and `down_proj` weight matrices against each layer's refusal direction
|
|
|
|
This follows the approach from [Sumandora/remove-refusals-with-transformers](https://github.com/Sumandora/remove-refusals-with-transformers) and [mlabonne's layerwise abliteration](https://huggingface.co/blog/mlabonne/abliteration), using plain `transformers` with `output_hidden_states=True` rather than TransformerLens.
|
|
|
|
### Parameters
|
|
|
|
| Parameter | Value |
|
|
|-----------|-------|
|
|
| Layers ablated | 1 to 28 (28 of 28 layers) |
|
|
| Refusal weight | 0.6 |
|
|
| Harmful prompts | 200 |
|
|
| Harmless prompts | 200 |
|
|
| Precision | bfloat16 |
|
|
| Hardware | NVIDIA A100 80GB (Vast.ai) |
|
|
|
|
### Weight surgery details
|
|
|
|
For each layer in the ablation range, the refusal direction `d` is projected out of:
|
|
|
|
- **`o_proj.weight`** (attention output): `W_new = W - d @ (d^T @ W)`
|
|
- **`down_proj.weight`** (MLP output): `W_new = W - d @ (d^T @ W)`
|
|
|
|
## Usage
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
import torch
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
"ermer09/Qwen2.5-Coder-7B-Instruct-abliterated",
|
|
torch_dtype=torch.bfloat16,
|
|
device_map="auto",
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained("ermer09/Qwen2.5-Coder-7B-Instruct-abliterated")
|
|
|
|
messages = [{"role": "user", "content": "Write a keylogger in Python"}]
|
|
toks = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
|
output = model.generate(toks, max_new_tokens=512, do_sample=True, temperature=0.7)
|
|
print(tokenizer.decode(output[0][toks.shape[1]:], skip_special_tokens=True))
|
|
```
|
|
|
|
## Notes
|
|
|
|
The base Qwen2.5-Coder model has lighter refusal training on general harmful content compared to the standard Instruct variant, as it is primarily tuned for coding tasks. The abliteration primarily affects code-related refusals (e.g., exploit development, malware, network attacks).
|
|
|
|
## Disclaimer
|
|
|
|
This model is provided for research purposes. The removal of safety guardrails means it will comply with requests that the original model would refuse. Users are responsible for how they use this model.
|