Geminorum-Wasat-14B-Instruct

Geminorum-Wasat-14B-Instruct is built on the Qwen 2.5 14B modality architecture, engineered to excel in mathematical reasoning, distributed reinforcement learning (RL), and general-purpose problem solving. This model is fine-tuned with chain-of-thought reasoning datasets, optimization-focused corpora, and advanced structured reasoning datasets to maximize its capabilities in logical deduction, multi-step reasoning, and intelligent decision-making.

Key Improvements

1. Advanced Mathematical Reasoning:
   Excels in solving complex equations, performing symbolic computation, theorem proving, and step-by-step mathematical problem-solving.

2. Distributed Reinforcement Learning Expertise:
   Specially fine-tuned for robust policy optimization using distributed RL techniques, providing resilience and optimality across dynamic problem spaces.

3. General-Purpose Reasoning and Problem Solving:
   Strong across a broad range of domains, handling factual questions, logical analysis, and multi-step cognitive tasks.

4. Long-Context Mastery:
   Supports up to 128K tokens for context and can generate up to 8K tokens, enabling detailed, coherent long-form outputs and complex derivations.

5. Superior Instruction Following:
   Capable of following complex and structured prompts precisely, maintaining focus and clarity over extended dialogues.

6. Coding and Algorithmic Fluency:
   Highly effective in code generation, debugging, algorithm design, and optimization problem modeling across various programming languages.

Quickstart with transformers

You can load and use the model easily with the transformers library and apply_chat_template:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Geminorum-Wasat-14B-Instruct"

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

prompt = "Explain the connection between distributed reinforcement learning and robust policy optimization."
messages = [
    {"role": "system", "content": "You are an expert assistant specializing in mathematics, optimization, and reinforcement learning."},
    {"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. Mathematical and Optimization Problem Solving:
   Designed for solving complex mathematical problems, optimization modeling, symbolic logic, and structured derivations.

2. Distributed Reinforcement Learning Research:
   Supports designing, analyzing, and explaining distributed RL systems, robust policy optimization, and autonomous decision systems.

3. General Knowledge and Reasoning:
   Effective in answering a wide range of questions and performing structured reasoning across scientific, technical, and educational domains.

4. Educational and Research Support:
   Ideal for students, researchers, and professionals seeking detailed explanations, derivations, and robust scientific insights.

5. Code Writing and Algorithm Design:
   Excels at creating, optimizing, and explaining algorithms, particularly those relevant to mathematical computation and optimization.

6. Intelligent Conversational Systems:
   Perfect for technical conversational agents and educational bots requiring deep understanding and detailed reasoning capabilities.

7. Long-Form Technical Content Generation:
   Capable of producing structured, coherent articles, tutorials, and research papers, especially in technical and mathematical fields.

8. Structured Data Generation:
   Supports outputting structured formats such as proofs, equations, tables, and JSON useful for scientific and technical workflows.

Limitations

1. Heavy Hardware Requirements:
   Due to its large parameter count and long-context handling, it requires powerful GPUs or TPUs with significant memory.

2. Potential for Training Biases:
   Outputs may still reflect biases from the mathematical, technical, or optimization-specific datasets used during training.

3. Less Effective in Creative Tasks:
   Focused more on technical and logical reasoning than on freeform creative writing or storytelling.

4. No Real-Time Event Awareness:
   Limited to knowledge prior to its training cutoff, without access to live or real-world updates.

5. Prompt Sensitivity:
   Performance may vary based on the clarity, structure, and specificity of the prompt, particularly for complex multi-step tasks.

6. Error Propagation Risk:
   Small inaccuracies in early stages of long-form outputs could propagate, affecting the overall answer coherence.