Taurus-Opus-7B

Taurus-Opus-7B is built upon the LLaMA (Large Language Model Meta AI) 7B architecture, optimized to provide advanced reasoning capabilities while maintaining efficiency. With 7 billion parameters, it strikes a balance between performance and computational resource requirements. The model has been fine-tuned with a focus on chain-of-thought (CoT) reasoning, leveraging specialized datasets to enhance its problem-solving abilities. Taurus-Opus-7B is designed for tasks requiring logical reasoning, detailed explanations, and multi-step problem-solving, making it ideal for applications such as instruction-following, text generation, and coding assistance.

Key Features and Improvements

1. Optimized Reasoning Capabilities:
   The model showcases significant improvements in context understanding, reasoning, and mathematical problem-solving through fine-tuning with long CoT datasets.

2. Enhanced Instruction Following:
   Taurus-Opus-7B excels in generating long, coherent outputs (up to 4K tokens), understanding structured data, and producing structured outputs like JSON.

3. Lightweight Efficiency:
   Its 7B parameter size makes it more resource-efficient compared to larger models while retaining high-quality performance for reasoning and content generation tasks.

4. Long-Context Support:
   Offers support for long contexts of up to 64K tokens, enabling the handling of large datasets or extended conversations.

5. Multilingual Proficiency:
   The model supports 20+ languages, including English, Spanish, French, German, Portuguese, Chinese, Japanese, and more, making it suitable for global applications.

Quickstart with transformers

Here’s a code snippet to load Taurus-Opus-7B using the transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Taurus-Opus-7B"

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

prompt = "Explain the importance of chain-of-thought reasoning in large language models."
messages = [
    {"role": "system", "content": "You are a helpful assistant with expertise in logical reasoning and problem-solving."},
    {"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:
   Taurus-Opus-7B is tailored for complex reasoning tasks, contextual understanding, and solving problems requiring logical deduction.

2. Mathematical Problem-Solving:
   Designed for advanced mathematical reasoning and calculations, making it valuable for education, research, and engineering tasks.

3. Code Assistance:
   Provides robust coding support, including writing, debugging, and optimizing code across multiple programming languages.

4. Data Analysis:
   Excels in analyzing structured data and generating structured outputs, aiding automation workflows and data-driven insights.

5. Multilingual Support:
   Facilitates applications such as multilingual chatbots, content generation, and translation in 20+ languages.

6. Extended Content Generation:
   Suitable for generating detailed reports, articles, and instructional guides, handling outputs up to 4K tokens.

Limitations

1. Hardware Requirements:
   While more efficient than larger models, Taurus-Opus-7B still requires high-memory GPUs or TPUs for optimal performance.

2. Language Quality Variations:
   Output quality may vary across supported languages, especially for less commonly used languages.

3. Creativity Limitations:
   The model may sometimes generate repetitive or inconsistent results in creative or highly subjective tasks.

4. Real-Time Knowledge Constraints:
   The model lacks awareness of events or knowledge updates beyond its training data.

5. Prompt Dependency:
   Results heavily depend on the specificity and clarity of input prompts, requiring well-structured queries for the best performance.

Open LLM Leaderboard Evaluation Results

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

Metric	Value (%)
Average	26.06
IFEval (0-Shot)	42.23
BBH (3-Shot)	34.23
MATH Lvl 5 (4-Shot)	22.73
GPQA (0-shot)	10.18
MuSR (0-shot)	14.22
MMLU-PRO (5-shot)	32.79