Calcium-Opus-14B-Elite3 is based on the Qwen 2.5 14B modality architecture, designed to enhance the reasoning capabilities of 14B-parameter models. These models have proven effective in context understanding, reasoning, and mathematical problem-solving. It has been fine-tuned using a long chain-of-thought reasoning model and specialized datasets, with a focus on chain-of-thought (CoT) reasoning for problem-solving. This model is optimized for tasks requiring logical reasoning, detailed explanations, and multi-step problem-solving, making it ideal for applications such as instruction-following, text generation, and complex reasoning tasks.

Key improvements include:

Enhanced Knowledge and Expertise: The model demonstrates significantly more knowledge and greatly improved capabilities in coding and mathematics, thanks to specialized expert models in these domains.
Improved Instruction Following: It shows significant advancements in following instructions, generating long texts (over 8K tokens), understanding structured data (e.g., tables), and producing structured outputs, especially in JSON format.
Better Adaptability: The model is more resilient to diverse system prompts, enabling enhanced role-playing implementations and condition-setting for chatbots.
Long-Context Support: It offers long-context support of up to 128K tokens and can generate up to 8K tokens in a single output.
Multilingual Proficiency: The model supports over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more.

Quickstart with transformers

Here provides a code snippet with apply_chat_template to show you how to load the tokenizer and model and how to generate contents.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Calcium-Opus-14B-Elite3"

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

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
    {"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

Reasoning and Context Understanding:
Designed to assist with complex reasoning tasks, contextual understanding, and solving problems requiring logical deduction and critical thinking.
Mathematical Problem-Solving:
Specialized for performing advanced mathematical reasoning and calculations, making it suitable for educational, scientific, and research-oriented applications.
Code Generation and Debugging:
Offers robust support for coding tasks, including writing, debugging, and optimizing code in various programming languages, ideal for developers and software engineers.
Structured Data Analysis:
Excels in processing and analyzing structured data, such as tables and JSON, and generating structured outputs, which is useful for data analysts and automation workflows.
Multilingual Applications:
Supports over 29 languages, making it versatile for global applications like multilingual chatbots, content generation, and translations.
Extended Content Generation:
Capable of generating long-form content (over 8K tokens), useful for writing reports, articles, and creating detailed instructional guides.

Limitations

Hardware Requirements:
Due to its 20B parameter size and support for long-context inputs, running the model requires significant computational resources, including high-memory GPUs or TPUs.
Potential Bias in Multilingual Outputs:
While it supports 29 languages, the quality and accuracy of outputs may vary depending on the language, especially for less-resourced languages.
Inconsistent Outputs for Creative Tasks:
The model may occasionally produce inconsistent or repetitive results in creative writing, storytelling, or highly subjective tasks.
Limited Real-World Awareness:
It lacks real-time knowledge of current events beyond its training cutoff, which may limit its ability to respond accurately to the latest information.
Error Propagation in Long-Text Outputs:
In generating long texts, minor errors in early outputs can sometimes propagate, reducing the overall coherence and accuracy of the response.
Dependency on High-Quality Prompts:
Performance may depend on the quality and specificity of the input prompt, requiring users to carefully design queries for optimal results.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	35.86
IFEval (0-Shot)	54.28
BBH (3-Shot)	47.07
MATH Lvl 5 (4-Shot)	29.38
GPQA (0-shot)	16.11
MuSR (0-shot)	20.13
MMLU-PRO (5-shot)	48.17