Update README.md

README.md

@@ -22,29 +22,29 @@ datasets:
 
 # Ophiuchi-Qwen3-14B-Instruct
 
-> Ophiuchi-Qwen3-14B-Instruct is built upon the Qwen3-14B architecture, featuring Qwen3ForCausalLM and optimized for high-performance mathematical reasoning, coding proficiency, and factual accuracy. This model extends the capabilities of large instruction-tuned models by integrating precision-focused datasets, advanced reasoning chains, and fine-grained instruction alignment.
+> Ophiuchi-Qwen3-14B-Instruct is built upon the Qwen3-14B architecture and uses the Qwen3ForCausalLM backbone. It is instruction-tuned to enhance capabilities in mathematical reasoning, code generation, and factual accuracy. By leveraging high-quality datasets and long-context architectures, this model is designed to excel in solving complex reasoning tasks and generating accurate, structured content across multiple domains.
 
-Key enhancements include:
+## Key Features
 
-1. Precision in Mathematical and Logical Reasoning
-   The model excels at solving complex mathematical problems, symbolic logic, and step-by-step derivations.
+1. Mathematical and Logical Reasoning
+   Fine-tuned to perform step-by-step reasoning, symbolic logic, and advanced mathematics, supporting educational and technical use cases.
 
-2. Robust Code Understanding and Generation
-   Tuned for code generation and interpretation across multiple programming languages with an emphasis on correctness and structure.
+2. Code Generation and Understanding
+   Optimized for writing, interpreting, and debugging code across various programming languages, including Python, JavaScript, and C++.
 
-3. Accurate Factual Reasoning
-   Designed to reduce hallucination, this model prioritizes reliable and verifiable knowledge across domains.
+3. Factual Integrity and Precision
+   Trained on curated and aligned datasets to enhance accuracy and reduce hallucination in fact-based tasks.
 
 4. Long-Context Support
-   Handles up to 128K tokens in the input with support for 8K-token outputs, enabling comprehensive multi-step responses.
+   Capable of handling up to 128K tokens as input with output generation up to 8K tokens, enabling detailed and comprehensive responses over extended sequences.
 
-5. Instruction-Tuned with High Fidelity
-   Responds precisely to multi-turn and nested instructions, maintaining structured and relevant outputs.
+5. Instruction-Tuned Alignment
+   Demonstrates a strong ability to follow multi-step instructions, maintain conversation context, and produce structured outputs across sessions.
 
-6. Multilingual Competence
-   Supports over 29 global languages, facilitating international deployment and multilingual content creation.
+6. Multilingual Proficiency
+   Supports over 29 languages including English, Chinese, French, Spanish, Arabic, Russian, Japanese, Korean, and others, enabling global communication and translation tasks.
 
-# Quickstart with Transformers
+## Quickstart with Transformers
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -58,7 +58,8 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-prompt = "What are the core principles behind large language model alignment?"
+prompt = "Explain the principles of alignment in large language models."
+
 messages = [
     {"role": "system", "content": "You are a highly capable assistant focused on reasoning, coding, and factual precision."},
     {"role": "user", "content": prompt}
@@ -84,19 +85,26 @@ response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
 print(response)
 ```
 
-# Intended Use
+## Intended Use
+
+* Mathematical and symbolic problem solving
+* Code generation and explanation
+* Structured response generation in JSON, Markdown, or table formats
+* Long-form technical writing and documentation
+* Factual question answering and fact-checking
+* Educational assistance across STEM domains
+* Multilingual conversation and translation tasks
+
+## Limitations
+
+* High computational requirements (A100/H100-class GPUs recommended)
+* May still produce hallucinated facts on edge cases or adversarial inputs
+* Sensitive to poorly structured or ambiguous prompts
+* Early-stage errors may propagate in long outputs
+* Less suitable for creative fiction or subjective narrative tasks
 
-* Advanced mathematical and logical problem-solving
-* Code generation and debugging in multiple languages
-* Technical documentation and instruction parsing
-* Structured outputs (e.g., tables, JSON, config files)
-* Factual Q\&A and explanation tasks
-* Educational support and multilingual tutoring
+## References
 
-# Limitations
+1. Saxton, D., Grefenstette, E., Hill, F., & Blunsom, P. (2019). Analysing Mathematical Reasoning Abilities of Neural Models. arXiv:1904.01557. [https://arxiv.org/pdf/1904.01557](https://arxiv.org/pdf/1904.01557)
 
-* Requires high-memory hardware for optimal performance
-* May still produce factual inaccuracies on niche or adversarial prompts
-* Sensitive to prompt phrasing; structured prompts yield better results
-* Long outputs may propagate initial reasoning flaws
-* Creative tasks (e.g., fiction writing) may be less consistent
+2. Chen, X., Zheng, S., & Liu, Z. (2023). YaRN: Efficient Context Window Extension of Large Language Models. arXiv:2309.00071. [https://arxiv.org/pdf/2309.00071](https://arxiv.org/pdf/2309.00071)