Finetuned from model : llama-3.2-3b-instruct-bnb-4bit
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
Deep-Reason-SMALL-V0
Overview
Deep-Reason-SMALL-V0 is a fine-tuned version of llama-3.2-3b-instruct, designed for advanced reasoning and thinking capabilities. It has been trained using Reasoning GRPO techniques and a custom dataset curated for enhancing logical inference, decision-making, and structured reasoning.
Built with Unsloth and Hugging Face’s TRL, this model is optimized for faster inference and superior logical performance.
The model is available in GGUF and 16 Bit format and has been quantized to different levels to support various hardware configurations.
Enhanced Reasoning: Fine-tuned using GRPO to improve problem-solving and structured thought processes.
Optimized for Thinking Tasks: Excels in logical, multi-step, and causal reasoning.
Structured XML Responses: Outputs are formatted using a structured reasoning-answer format for easy parsing. Outputs are formatted using structured <reasoning>...</think> and <answer>...</answer> sections for easy parsing.
Efficient Deployment: Available in GGUF format for local AI deployments on consumer hardware.
Response Format & Parsing Instructions
Deep-Reason-SMALL-V0 follows a structured response format with designated XML-like tags for easy parsing. The XML responses will include tokens such as <reasoning>...</reasoning> and <answer>...</answer>. Users must extract the tokens accordingly when using programmatically. This ensures clarity and traceability in decision-making.
Ideal Configuration for using the GGUF Models
temperature = 0.8
top_p = 0.95
max_tokens = 1024
SYSTEM_PROMPT = """
Respond in the following format:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
"""
Use Cases
Deep-Reason-SMALL-V0 is best suited for:
Conversational AI – Improving chatbot and AI assistant reasoning.
AI Research – Studying logical thought modeling in AI.
Automated Decision Making – Use in AI-powered business intelligence systems.
Education & Tutoring – Helping students and professionals with structured learning.
Legal & Financial Analysis – Generating step-by-step arguments for case studies.
Limitations & Considerations
May require further fine-tuning for domain-specific logic.
Not a factual knowledge base – Focused on reasoning, not general knowledge retrieval.
Potential biases – Results depend on training data.
Computational Trade-off – Reasoning performance comes at the cost of slightly longer inference times.
License
This model is released under a permissible license.
Acknowledgments
Special thanks to the Unsloth team for providing an optimized training pipeline for LLaMA models.
