Update Readme.md

Readme.md

@@ -13,49 +13,39 @@ tags:
 - instruction-tuned
 - pytorch
 ---
-
-# **Supertron1-8B: A Capable Instruction-Tuned Language Model**
-
+# **Supertron1-8B: A Capable, Efficient Instruction-Tuned Language Model**
 ## **Model Description**
-
-**Supertron1-8B** is an instruction-tuned language model built on top of Qwen3-8B. It is designed to be a strong general-purpose assistant for reasoning, coding, math, and everyday chat while keeping the workflow simple for local use and Hugging Face deployment.
+**Supertron1-8B** is an instruction-tuned language model built on top of Qwen3-8B-Base. Designed to be a **reliable, efficient daily driver**, it delivers strong performance across math, coding, reasoning, and general conversation while remaining fast enough to run on consumer hardware with a capable GPU.
 
 * **Developed by:** Surpem
 * **Model type:** Causal Language Model
 * **Architecture:** Dense Transformer, 8B parameters
-* **Fine-tuned from:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)
+* **Fine-tuned from:** [Qwen/Qwen3-8B-Base](https://huggingface.co/Qwen/Qwen3-8B-Base)
+* **Fine-tuning method:** LoRA (r=16, alpha=32, all-linear targets)
 * **License:** Apache 2.0
 
 ---
 
-## **Highlights**
+## **Capabilities**
 
-Supertron1-8B is intended as a balanced model for:
+### **Reasoning**
+Supertron1-8B was trained on long-form chain-of-thought reasoning traces, making it capable of breaking down complex multi-step problems clearly and methodically. It thinks through problems before answering rather than jumping to conclusions, resulting in more reliable and explainable outputs.
 
-- Instruction following
-- Coding help
-- Mathematical reasoning
-- General conversation
+### **Math**
+With dedicated training on competition-style math problems and step-by-step solutions, the model handles everything from algebra and calculus to word problems with structured, verifiable working. It consistently shows its reasoning rather than just producing a final answer.
 
-With its 8B-class size, it offers a stronger capability tier than smaller models while remaining practical for many local inference setups.
+### **Coding**
+Supertron1-8B can write, debug, and explain code across popular languages including Python, JavaScript, C++, and more. Trained on filtered, high-quality coding instruction data, it understands not just syntax but software design patterns, algorithmic thinking, and best practices.
 
----
+### **Science & General Knowledge**
+Broad instruction tuning across science, STEM, and general knowledge domains means the model can hold detailed technical conversations, explain difficult concepts clearly, and assist with research, writing, and analysis tasks.
 
-## **Model Details**
-
-Based on the included model configuration:
-
-- **Context length:** 32,768 tokens
-- **Hidden size:** 4096
-- **Layers:** 36
-- **Attention heads:** 32
-- **KV heads:** 8
-- **Default dtype:** bfloat16
+### **Instruction Following**
+The model is highly responsive to natural language instructions. Whether you need concise answers, detailed explanations, structured output, or creative writing, Supertron1-8B adapts to the format and tone you ask for without needing complex prompting tricks.
 
 ---
 
 ## **Get Started**
-
 ```python
 from transformers import AutoTokenizer, AutoModelForCausalLM
 import torch
@@ -75,26 +65,34 @@ messages = [
 
 text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
 inputs = tokenizer(text, return_tensors="pt").to(model.device)
-
 outputs = model.generate(**inputs, max_new_tokens=512)
 print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
 ```
 
 ---
 
-## **License**
+## **Hardware Requirements**
+| Precision | Min VRAM | Recommended |
+|---|---|---|
+| bfloat16 | 18 GB | 24 GB (RTX 3090/4090) |
+| 4-bit quantized | 8 GB | 12 GB (RTX 3060/4070) |
+
+For 4-bit quantized inference:
+```python
+from transformers import BitsAndBytesConfig
 
-This model is released under the Apache License 2.0. See the [LICENSE](./LICENSE) file for details.
+bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
+model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map="auto")
+```
 
 ---
 
 ## **Citation**
-
 ```bibtex
-@misc{surpem2026supertron1,
-      title={Supertron1-8B - Instruction-Tuned Language Model},
+@misc{surpem2026supertron1-8b,
+      title={Supertron1-8B — Efficient Instruction-Tuned Language Model},
       author={Surpem},
       year={2026},
       url={https://huggingface.co/surpem/supertron1-8b},
 }
-```
+```