metadata

tags:
  - long-cot-reasoning
  - transformers
  - mamba2
  - llms
  - chain-of-thought
license: apache-2.0
language:
  - en
datasets:
  - Daemontatox/LongCOT-Reason
  - Daemontatox/alpaca_reasoning_COT
base_model:
  - Qwen/Qwen2.5-14B-Instruct
pipeline_tag: text-generation
library_name: transformers
model-index:
  - name: Sphinx2.0
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: wis-k/instruction-following-eval
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 71.23
            name: averaged accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: SaylorTwift/bbh
          split: test
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 49.4
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: lighteval/MATH-Hard
          split: test
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 2.72
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 5.82
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 13.05
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 46.49
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FSphinx2.0
          name: Open LLM Leaderboard

Sphinx: The Apex of Logical Deduction and Chain-of-Thought Reasoning

Developed by: Daemontatox
License: Apache-2.0
Base Model: Fine-tuned from unsloth/qwen2.5-14b-instruct-bnb-4bit
Accelerated by: Unsloth Framework
TRL-Optimized: Integrated with Huggingface's TRL library for enhanced performance in logical reasoning.

Unveiling Sphinx: Master of Reasoned Thought

Sphinx is a cutting-edge Long Chain-of-Thought (CoT) reasoning model meticulously crafted to unravel complex challenges requiring rigorous logical analysis. Built upon the robust foundation of the Qwen2.5 architecture, Sphinx excels at constructing coherent, step-by-step thought processes, providing unparalleled insight into its reasoning and ensuring clarity in its conclusions.

"Where complexity yields to logical clarity."

Core Strengths: Reasoning, Logic, and CoT

Unrivaled Chain-of-Thought (CoT) Mastery: Engineered for dissecting intricate problems, Sphinx meticulously constructs each step of its reasoning, offering a transparent and verifiable pathway to the solution.
Deep Logical Reasoning Capabilities: Sphinx is adept at navigating complex logical structures, drawing valid inferences and forming sound conclusions through multi-layered analysis.
Exceptional Reasoning Fidelity: Fine-tuned to maintain the highest standards of logical consistency, Sphinx delivers outputs that are not only correct but also demonstrably well-reasoned.
Efficient Long-Context Reasoning: Leveraging the power of Unsloth, Sphinx processes extensive information efficiently, maintaining logical coherence across extended reasoning chains.
Explainable AI through Transparent Logic: Sphinx's inherent CoT approach provides explicit and understandable reasoning, making its decision-making process transparent and trustworthy.

Model Architecture and Fine-tuning for Logical Prowess

Architectural Foundation

Base Model: Qwen2.5-14B - Renowned for its strong general language understanding, forming a solid basis for specialized reasoning.
Parameters: 14 billion - Providing the capacity to model intricate reasoning patterns.
Quantization: 4-bit precision using BitsAndBytes (bnb) - Optimizing for accessibility without sacrificing logical reasoning accuracy.
Extended Reasoning Window: Supports inputs up to 16k tokens, crucial for accommodating the detailed context required for complex logical deductions.

Training Methodology: Honing Logical Acumen

Frameworks: Huggingface Transformers + TRL + Unsloth - A powerful combination for efficient training and reinforcement learning.
Data Sources: A meticulously curated collection of datasets specifically designed to challenge and refine logical reasoning skills, encompassing academic, legal, and formal logic domains.
Optimization Strategies:
- LoRA (Low-Rank Adaptation): Enabling parameter-efficient fine-tuning, focusing on adapting the model for superior logical inference.
- Reinforcement Learning from Human Feedback (RLHF): Guiding the model towards generating more logically sound and human-aligned reasoning steps.

Sphinx's Reasoning Toolkit: Capabilities in Action

Masterful Long-CoT Generation: Deconstructs and conquers multi-layered problems by constructing detailed, logically interconnected reasoning sequences.
Explanatory Power through Logic: Provides clear, step-by-step logical derivations for its outputs, enhancing trust and understanding.
Adaptable Logical Framework: Easily tailored to specialized reasoning tasks through targeted fine-tuning, enabling application in diverse logical domains.

Unlocking Potential: Applications Driven by Logic

Advanced Academic Research: Generating in-depth, logically structured analyses for complex scientific and philosophical inquiries.
Robust Legal Reasoning Assistance: Constructing and articulating multi-step legal arguments with precision and logical rigor.
Transformative STEM Education: Guiding learners through intricate mathematical and logical problems with clear, step-by-step explanations.
Transparent Cognitive AI Systems: Powering AI systems where explainability and logical justification are paramount for decision-making.# Open LLM Leaderboard Evaluation Results Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	31.45
IFEval (0-Shot)	71.23
BBH (3-Shot)	49.40
MATH Lvl 5 (4-Shot)	2.72
GPQA (0-shot)	5.82
MuSR (0-shot)	13.05
MMLU-PRO (5-shot)	46.49

Open LLM Leaderboard Evaluation Results

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

Metric	Value (%)
Average	31.45
IFEval (0-Shot)	71.23
BBH (3-Shot)	49.40
MATH Lvl 5 (4-Shot)	2.72
GPQA (0-shot)	5.82
MuSR (0-shot)	13.05
MMLU-PRO (5-shot)	46.49