Calcium-Opus-20B-v1

Calcium-Opus-20B-v1 is based on the Qwen 2.5 modality architecture, designed to enrich the reasoning capabilities of 20B-parameter models. These models have proven highly effective for context understanding, reasoning, and mathematical problem-solving.

Key improvements include:

1. 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.
2. 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.
3. Better Adaptability: The model is more resilient to diverse system prompts, enabling enhanced role-playing implementations and condition-setting for chatbots.
4. Long-Context Support: It offers long-context support of up to 128K tokens and can generate up to 8K tokens in a single output.
5. 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.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Calcium-Opus-20B-v1"

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

1. Reasoning and Context Understanding:
   Designed to assist with complex reasoning tasks, contextual understanding, and solving problems requiring logical deduction and critical thinking.

2. Mathematical Problem-Solving:
   Specialized for performing advanced mathematical reasoning and calculations, making it suitable for educational, scientific, and research-oriented applications.

3. 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.

4. 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.

5. Multilingual Applications:
   Supports over 29 languages, making it versatile for global applications like multilingual chatbots, content generation, and translations.

6. Extended Content Generation:
   Capable of generating long-form content (over 8K tokens), useful for writing reports, articles, and creating detailed instructional guides.

7. Interactive Role-Playing and Chatbots:
   Enhanced capabilities for role-playing and condition-setting, making it ideal for interactive chatbots, virtual assistants, and entertainment purposes.

8. Large-Context Tasks:
   With a context window of up to 128K tokens, it is ideal for analyzing or generating large documents, books, or datasets in a single session.

Limitations

1. 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.

2. 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.

3. Inconsistent Outputs for Creative Tasks:
   The model may occasionally produce inconsistent or repetitive results in creative writing, storytelling, or highly subjective tasks.

4. 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.

5. 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.

6. 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.

7. Sensitivity to Adversarial Inputs:
   The model may struggle with adversarial or ambiguous inputs, leading to incorrect or irrelevant outputs.

8. Ethical and Safety Concerns:
   Potential misuse in generating misleading, harmful, or offensive content remains a concern, and guardrails must be implemented to ensure responsible use.

Open LLM Leaderboard Evaluation Results

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

Metric	Value (%)
Average	26.85
IFEval (0-Shot)	30.93
BBH (3-Shot)	41.81
MATH Lvl 5 (4-Shot)	11.03
GPQA (0-shot)	13.76
MuSR (0-shot)	22.09
MMLU-PRO (5-shot)	41.49