| --- |
| license: other |
| base_model: Qwen/Qwen3-4B-Instruct-2507 |
| tags: |
| - qwen |
| - qwen3 |
| - fine-tuned |
| - safety |
| - instruct |
| - axion |
| --- |
| |
| # AdvRahul/Axion-4B |
|
|
| **A safety-enhanced version of Qwen3-4B-Instruct, optimized for reliable and responsible AI applications. 🛡️** |
|
|
| `Axion-4B` is a fine-tuned version of the powerful `Qwen/Qwen3-4B-Instruct-2507` model. The primary enhancement in this version is its **robust safety alignment**, making it a more dependable choice for production environments and user-facing applications. |
|
|
| ## 🚀 Model Details |
|
|
| * **Model Creator:** AdvRahul |
| * **Base Model:** [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) |
| * **Fine-tuning Focus:** Enhanced Safety & Harmlessness via Red-Teaming |
| * **Architecture:** Qwen3 |
| * **Context Length:** 262,144 tokens |
| * **License:** Based on the Tongyi Qianwen LICENSE of the original model. |
|
|
| --- |
|
|
| ## 📝 Model Description |
|
|
| ### Enhanced for Safety |
|
|
| The core purpose of `Axion-4B` is to provide a safer alternative for developers. The base model underwent extensive **red-team testing using advanced protocols** to significantly minimize the generation of harmful, biased, or inappropriate content. |
|
|
| ### Powerful Core Capabilities |
|
|
| While adding a crucial safety layer, `Axion-4B` retains the exceptional capabilities of its base model, including: |
| * **Strong Logical Reasoning:** Excels at complex problems in math, science, and logic. |
| * **Advanced Instruction Following:** Reliably adheres to user commands and constraints. |
| * **Multi-lingual Knowledge:** Covers a wide range of languages and cultural contexts. |
| * **Massive 256K Context Window:** Capable of understanding and processing very long documents. |
| * **Excellent Coding & Tool Use:** Proficient in code generation and agentic tasks. |
|
|
| --- |
|
|
| ## 💻 Quickstart |
|
|
| You can use this model directly with the `transformers` library (version 4.51.0 or newer is recommended). |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| # IMPORTANT: Use the model name for this repository |
| model_name = "AdvRahul/Axion-4B" |
| |
| # Load the tokenizer and the model |
| tokenizer = AutoTokenizer.from_pretrained(model_name) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_name, |
| torch_dtype="auto", |
| device_map="auto" |
| ) |
| |
| # Prepare the model input |
| prompt = "Give me a short introduction to large language models and their safety considerations." |
| messages = [ |
| {"role": "user", "content": prompt} |
| ] |
| text = tokenizer.apply_chat_template( |
| messages, |
| tokenize=False, |
| add_generation_prompt=True, |
| ) |
| model_inputs = tokenizer([text], return_tensors="pt").to(model.device) |
| |
| # Generate text |
| generated_ids = model.generate( |
| **model_inputs, |
| max_new_tokens=512 # Limiting for a concise example |
| ) |
| output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() |
| |
| content = tokenizer.decode(output_ids, skip_special_tokens=True) |
| print("Response:", content) |
| |
| Optimized Deployment |
| For high-throughput, production-ready deployment, you can use frameworks like vLLM or SGLang to serve the model via an OpenAI-compatible API. |
| |
| vLLM: |
| |
| Bash |
| |
| vllm serve AdvRahul/Axion-4B --max-model-len 262144 |
| SGLang: |
| |
| Bash |
| |
| python -m sglang.launch_server --model-path AdvRahul/Axion-4B --context-length 262144 |
| Note: If you encounter out-of-memory (OOM) issues, consider reducing the max context length (e.g., --max-model-len 32768). |
| |
| ⚠️ Ethical Considerations and Limitations |
| This model was fine-tuned with the explicit goal of improving safety and reducing harmful outputs. However, no AI model is completely immune to risks. |
| |
| No Guarantees: While the safety alignment is significantly improved, it does not guarantee perfectly harmless outputs in all scenarios. |
| |
| Inherited Biases: The model may still reflect biases present in the vast amount of data used to train its base model. |
| |
| Factual Accuracy: Always fact-check critical information, as the model can generate plausible but incorrect statements. |
| |
| Best Practice: It is strongly recommended that developers implement their own content moderation filters and safety guardrails as part of a comprehensive, defense-in-depth strategy. Thoroughly evaluate the model's performance and safety for your specific use case before deploying it to a live audience. |
| |
