Upload README.md with huggingface_hub

README.md

@@ -21,7 +21,30 @@ pipeline_tag: image-text-to-text
 
 ## Model Description
 
-LogicFlow-gemma-3-27b-thinking is a fine-tuned **multimodal** version of [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it) that has been specifically optimized for logical reasoning, step-by-step thinking, and mathematical problem-solving with both text and image inputs. This model has been trained using LoRA (Low-Rank Adaptation) technique and then merged with the base model for optimal performance.
+LogicFlow-gemma-3-27b-thinking is an advanced **multimodal reasoning model** built upon [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it), specifically designed to excel at complex logical reasoning, mathematical problem-solving, and step-by-step analytical thinking. This model represents a significant advancement in AI reasoning capabilities, achieved through careful fine-tuning on three specialized, high-quality datasets using LoRA (Low-Rank Adaptation) technique.
+
+### Training Dataset Foundation
+
+Our model has been meticulously trained on three cutting-edge datasets, each contributing unique reasoning capabilities:
+
+#### 🧠 **OpenO1-SFT Dataset**
+- **Purpose**: Supervised fine-tuning for advanced reasoning patterns
+- **Content**: High-quality reasoning demonstrations with explicit thought processes
+- **Impact**: Enables the model to break down complex problems systematically and show transparent reasoning chains
+
+#### 💭 **Open-Thoughts Dataset** 
+- **Purpose**: Step-by-step thinking process modeling
+- **Content**: Detailed internal monologues and reasoning progressions for various problem types
+- **Impact**: Teaches the model to externalize its thinking process, making reasoning transparent and verifiable
+
+#### 🔢 **OpenR1-Math Dataset**
+- **Purpose**: Mathematical reasoning and problem-solving specialization  
+- **Content**: Comprehensive mathematical problems with detailed solution methodologies
+- **Impact**: Significantly enhances performance on mathematical reasoning tasks, from basic arithmetic to advanced competition-level problems
+
+### Key Innovations
+
+This unique combination of datasets creates a model that not only provides correct answers but also demonstrates **how** it arrives at those answers, making it particularly valuable for educational applications, research, and any scenario requiring explainable AI reasoning.
 
 The model demonstrates enhanced capabilities in:
 - **🧠 Logical Reasoning**: Improved ability to work through complex logical problems step by step
@@ -48,10 +71,15 @@ The model demonstrates enhanced capabilities in:
 ## Training Details
 
 ### Training Data
-The model was fine-tuned on a combination of high-quality datasets:
-- **openo1_sft**: Supervised fine-tuning data for reasoning
-- **open_thoughts**: Dataset focused on step-by-step thinking processes
-- **open_r1_math**: Mathematical reasoning and problem-solving dataset
+The model was fine-tuned on three carefully selected, high-quality datasets that form the foundation of its exceptional reasoning capabilities:
+
+- **🧠 OpenO1-SFT**: Advanced supervised fine-tuning dataset containing high-quality reasoning demonstrations with explicit thought processes, enabling systematic problem breakdown and transparent reasoning chains
+
+- **💭 Open-Thoughts**: Specialized dataset focused on step-by-step thinking processes, featuring detailed internal monologues and reasoning progressions that teach the model to externalize and structure its thinking
+
+- **🔢 OpenR1-Math**: Comprehensive mathematical reasoning dataset with detailed solution methodologies, significantly enhancing performance from basic arithmetic to advanced competition-level mathematical problems
+
+This synergistic combination creates a model that excels not only at providing accurate answers but also at demonstrating clear, verifiable reasoning processes.
 
 ### Training Configuration
 
@@ -88,12 +116,12 @@ The model was fine-tuned on a combination of high-quality datasets:
 - **Freeze Multi-modal Projector**: true
 
 #### Special Features
-- **Enable Thinking**: true (enhanced reasoning capability)
-- **Template**: gemma
-- **Trust Remote Code**: true
-- **Preprocessing Workers**: 16
-- **Save Steps**: 100
-- **Logging Steps**: 5
+- **Enable Thinking**: true (**Critical** - Activates advanced Chain-of-Thought reasoning from OpenO1-SFT and Open-Thoughts datasets)
+- **Template**: gemma (Optimized for multimodal reasoning tasks)
+- **Trust Remote Code**: true (Required for advanced vision capabilities)
+- **Preprocessing Workers**: 16 (Optimized for multimodal data processing)
+- **Save Steps**: 100 (Frequent checkpointing for training stability)
+- **Logging Steps**: 5 (Detailed training monitoring)
 
 ### Training Results
 
@@ -122,8 +150,6 @@ The loss curve demonstrates stable convergence with the final training loss reac
 
 ### Comprehensive Evaluation Results
 
-Following established AI benchmarking best practices [(Domino AI, 2020)](https://domino.ai/blog/benchmarking-predictive-models), we conducted systematic evaluations across multiple domains to assess both predictive performance and operational characteristics. As emphasized by [(Cohere, 2025)](https://cohere.com/blog/ai-benchmarks-for-business), effective AI evaluation requires testing beyond simple accuracy metrics to capture real-world complexity and business needs.
-
 | **Benchmark** | **Metric** | **Base Gemma-3-27B-IT** | **LogicFlow-gemma-3-27b-thinking** | **Improvement** |
 |---------------|------------|--------------------------|-------------------------------------|-----------------|
 | **📊 Mathematical Reasoning** |
@@ -294,7 +320,6 @@ This model uses the standard Gemma 3 multimodal chat template with optimized for
 
 #### Text-only Chat
 ```python
-# Assuming model_name = "RekklesAI/LogicFlow-gemma-3-27b-thinking" is already defined
 messages = [
     {"role": "system", "content": "You are a helpful AI assistant specialized in logical reasoning and mathematics."},
     {"role": "user", "content": "Explain the reasoning behind the Pythagorean theorem and provide a step-by-step proof."}
@@ -319,7 +344,6 @@ print(response)
 ```python
 from PIL import Image
 
-# Assuming model and processor for "RekklesAI/LogicFlow-gemma-3-27b-thinking" are already loaded
 # Load an image
 image = Image.open("path/to/your/image.jpg")
 
@@ -502,7 +526,7 @@ For full reproducibility, here is the complete training configuration used:
 ```yaml
 bf16: true
 cutoff_len: 2048
-dataset: openo1_sft,open_thoughts,open_r1_math
+dataset: openo1_sft,open_thoughts,open_r1_math  # Three specialized reasoning datasets
 dataset_dir: data
 ddp_timeout: 180000000
 do_train: true