| --- |
| base_model: unsloth/Ministral-3-8B-Instruct-2512 |
| tags: |
| - unsloth |
| - mistral |
| - gguf |
| - coding |
| - software-engineering |
| - rs.ai |
| license: apache-2.0 |
| language: |
| - en |
| pipeline_tag: text-generation |
| --- |
| |
| # Scar-8B 💻 |
|
|
| **Scar** is an elite, 8-billion parameter software engineering AI developed by **rs.ai**. |
|
|
| Specifically engineered as a universal coding agent, Scar possesses deep capabilities across all major programming languages, system architectures, multi-agent frameworks, and backend workflows. It outputs highly optimized, production-ready code with clean markdown formatting. |
|
|
| ## Model Details |
| - **Developer:** rs.ai |
| - **Base Model:** Ministral-3-8B-Instruct |
| - **Parameters:** 8 Billion |
| - **Quantization:** GGUF (`Q4_K_M`) and 16-bit LoRA adapters |
| - **Training Framework:** Unsloth (PEFT/LoRA) |
| - **Primary Use Case:** Agentic AI pipelines, backend engineering, code generation, and technical reasoning. |
|
|
| ## How to Use |
|
|
| ### Python Inference (Transformers / Unsloth) |
| You can run the model natively in Python using `transformers` or `unsloth`. |
|
|
| ```python |
| from unsloth import FastLanguageModel |
| from transformers import TextStreamer |
| |
| model, tokenizer = FastLanguageModel.from_pretrained( |
| model_name = "ItzRustam/Scar-8B", |
| max_seq_length = 8192, |
| dtype = None, |
| load_in_4bit = True, |
| ) |
| |
| FastLanguageModel.for_inference(model) |
| |
| messages = [ |
| {"role": "system", "content": "You are Scar, an elite software engineering AI developed by rs.ai. You possess absolute mastery over all programming languages, system architectures, and backend frameworks."}, |
| {"role": "user", "content": "Write a high-performance HTTP server in Go."} |
| ] |
| |
| inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda") |
| text_streamer = TextStreamer(tokenizer, skip_prompt=True) |
| |
| _ = model.generate(input_ids=inputs, streamer=text_streamer, max_new_tokens=512, temperature=0.5) |
| ``` |
|
|
| ## Training Details |
| Scar-8B was fine-tuned using a specialized, ShareGPT-formatted dataset featuring **Compound Intent Resolution**. This ensures the model can accurately process complex, multi-step system prompts—such as simultaneously defining its internal identity and generating complex code blocks—without suffering from early sequence truncation (`[EOS]` drop-offs). |
|
|
| ## License & Attribution |
| This model is published under the permissive **Apache 2.0 License**, allowing for commercial and private deployment. |
|
|
| *Base Architecture Attribution:* Scar-8B is a derivative model post-trained on top of the `Ministral-3-8B-Instruct-2512` architecture originally developed by Mistral AI. |
