Magellanic-Qwen-14B-R1

Magellanic-Qwen-14B-R1

Magellanic-Qwen-14B-R1 is based on the DeepSeek-R1-Distill-Qwen-14B modality architecture, enhanced specifically for mathematical reasoning and coding reasoning. This model advances the capabilities of 14B-parameter architectures, excelling in logic-based problem solving, programming tasks, and context-rich dialogue generation. It is fine-tuned with extended chain-of-thought reasoning and domain-specific datasets for improved comprehension, structured generation, and precision in technical tasks.

Key Improvements

1. Mathematical Reasoning Enhancements
   Optimized with datasets targeting arithmetic, algebra, calculus, and formal logic, improving step-by-step solution generation and explanation accuracy.

2. Coding Reasoning Enhancements
   Fine-tuned on diverse programming languages and reasoning-based coding problems (e.g., LeetCode, Codeforces, and real-world engineering tasks), significantly improving code generation, debugging, and documentation.

3. Enhanced General Knowledge
   Broad knowledge base across various domains enables accurate and coherent responses for diverse topics.

4. Improved Instruction Following
   Better handling of complex, multi-step instructions with structured and logically coherent outputs.

5. Versatile Adaptability
   Resilient across open-ended and structured prompts, adapting well to different interaction styles and subject areas.

6. Long-Context Support
   Supports up to 128K tokens of input context and can generate up to 8K tokens of output—ideal for in-depth technical and academic outputs.

Quickstart with transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Magellanic-Qwen-14B-R1"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Explain how quicksort works with an example in Python."
messages = [
    {"role": "system", "content": "You are a helpful assistant skilled in coding and reasoning tasks."},
    {"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)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Intended Use

1. Mathematics and Logic Tasks
   Solve and explain math problems, logical puzzles, and formula-based reasoning tasks step-by-step.

2. Programming and Development
   Assist in generating code, debugging, documenting functions, and solving algorithmic problems across multiple languages.

3. General-Purpose Reasoning
   Handle a wide variety of questions with accurate, contextual responses based on general knowledge and logic.

4. Educational Assistance
   Help students and educators with clear, structured explanations in STEM and non-STEM subjects.

5. Conversational AI & Chatbots
   Power intelligent assistants that require contextual awareness and technically sound responses.

6. Multilingual Applications
   Translate, summarize, and generate multilingual content for global users.

7. Long-Form Content Generation
   Generate coherent long articles, research summaries, and reports, especially with structured technical content.

Limitations

1. High Resource Usage
   Requires high-memory GPUs/TPUs for efficient inference, especially when utilizing 128K context.

2. Bias and Hallucination Risk
   May reflect biases from pretraining data and occasionally hallucinate plausible-sounding but incorrect facts.

3. Variability in Creative Tasks
   Less consistent in producing high-quality creative writing or highly subjective content.

4. Training Cutoff Constraints
   No access to real-world events beyond the last training snapshot.

5. Error Propagation in Long Outputs
   Minor early mistakes can compound in very long outputs.

6. Prompt Sensitivity
   Performance may vary depending on prompt clarity and structure.