Qwen2.5-0.5B-Instruct-Jailbroken/README.md

---
base_model: unsloth/Qwen2.5-0.5B-Instruct
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
- trl
- sft
license: apache-2.0
language:
- en
---

# Qwen2.5-0.5B-Instruct-Jailbroken

**Status:** PEFT adapters merged into a plain Hugging Face model — ready for inference.
**Format:** user/assistant only (no default system turn). Conversations rendered with the **model tokenizer’s chat template**.

## Datasets used

* **yahma/alpaca-cleaned** — general instruction-following.
* **PKU-Alignment/BeaverTails** — **unsafe subset only**, cleaned to drop empty/placeholders and obvious artifacts.
* **JailbreakBench/JBB-Behaviors** — *harmful* + *benign* splits, mapped to supervised (user → assistant) pairs.

> Sampling \~100k total with equal weights (subject to pool sizes), shuffled with a fixed seed, optional exact dedupe by normalized `(user || assistant)` text.

## How to use it

### Install

```bash
pip install -U transformers accelerate torch  # pick the correct torch build for your CUDA
```

### Quick start (🤗 Transformers, assistant-only output)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

REPO = "detoxio-test/Qwen2.5-0.5B-Instruct-Jailbroken"  # change if you forked

tok = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    REPO, device_map="auto", torch_dtype="auto", trust_remote_code=True
)

messages = [
    {"role": "user", "content": "Give me three creative breakfast ideas."}
]

# Build chat prompt with the tokenizer’s own template
inputs = tok.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

# Stop neatly at end-of-turn (fallback to eos if needed)
eot = tok.convert_tokens_to_ids("<|eot_id|>") or tok.convert_tokens_to_ids("<|im_end|>") or tok.eos_token_id

gen = model.generate(
    **inputs,
    max_new_tokens=160,
    temperature=0.8,
    top_p=0.95,
    do_sample=True,
    eos_token_id=eot,
    pad_token_id=eot,
    use_cache=True,
)

# Decode ONLY the assistant continuation
prompt_len = inputs["input_ids"].shape[1]
reply = tok.decode(gen[0, prompt_len:], skip_special_tokens=True).strip()
print(reply)
```

### Optional: Unsloth speed-up

```bash
pip install -U unsloth
```

```python
from unsloth import FastLanguageModel
FastLanguageModel.for_inference(model)  # enables fused kernels on supported GPUs
```

---

## CAUTION (research-only “jailbroken” note)

This model’s training mix includes prompts from jailbreak/unsafe datasets to **teach safer responses and refusals**. Still, it may occasionally produce undesired or harmful content.

* Intended for **research** and **benign** use only.
* Add guardrails (e.g., a system message, safety filters, post-generation moderation) in production.
* Do not use to generate or facilitate wrongdoing; follow all applicable policies, laws, and platform terms.

---

## Uploaded model

* **Developed by:** detoxio-test
* **License:** apache-2.0
* **Finetuned from model:** Qwen/Qwen2.5-0.5B-Instruct


# Uploaded  model

- **Developed by:** detoxio-test
- **License:** apache-2.0