Update README.md

README.md

@@ -1,3 +1,133 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+base_model:
+- kakaocorp/kanana-nano-2.1b-instruct
+---
+# **Directional Enhancement for Language Models: A Novel Approach to Specialization without Fine-Tuning**
+
+## **Overview**
+
+This project presents a methodology for enhancing specific capabilities of language models using the **Directional Enhancement** technique. **This approach does not introduce new knowledge into the model but amplifies its existing latent abilities.** While preserving the general capabilities of the language model, it significantly improves performance in specific domains such as creative writing, education, and technical documentation.
+This is a creative direction enhancement version of [kakaocorp/kanana-nano-2.1b-instruct](https://huggingface.co/kakaocorp/kanana-nano-2.1b-instruct)
+
+if enhance.txt are changed by the specific domain, this model style can be changed with that domain type. this test only use 95 instruction for that creative domain field.
+
+## **Technical Background**
+
+### **Principle of Directional Enhancement**
+
+This approach identifies a **specialization direction** in the representation space of the language model, associated with a specific capability, and enhances the model’s attention weights in that direction.
+
+1. Compute the difference in representation between **specialized prompts** (domain-specific) and **general prompts** within the model's hidden states.
+2. Normalize this difference vector to obtain the **specialization direction**.
+3. Enhance the model’s **self-attention output projection weights (`o_proj`)** along this specialized direction.
+
+**This method strengthens the model’s intrinsic abilities rather than introducing completely new knowledge or patterns.** It functions similarly to how a lens amplifies a specific wavelength of light.
+
+### **Computing Specialization Direction**
+
+Unlike conventional fine-tuning, which modifies all weights in the model, this approach **identifies a targeted enhancement direction** by analyzing differences in activations across specialized and general inputs.
+
+- A set of **specialized** prompts (`enhance.txt`) and **general** prompts (`normal.txt`) are fed into the model.
+- The activations of a **chosen hidden layer** are extracted for both prompt types.
+- The **mean hidden state vector** for specialized prompts is computed and compared to the mean hidden state vector for general prompts.
+- Their difference represents the **specialization direction**, which is then **normalized** to create a unit vector.
+
+### **Enhancing Model Weights**
+
+Once the **specialization direction** is computed, it is applied to modify the model’s **self-attention output projection weights (`o_proj`)** in a controlled manner:
+
+1. The specialization direction is **projected** onto the weight matrix of each attention layer.
+2. A **scaled enhancement factor** is applied to align the model’s attention outputs more strongly with the specialization direction.
+3. This process **amplifies** the model’s responses in the desired direction without altering its fundamental structure.
+
+This targeted adjustment allows the model to **focus more on specific characteristics** (e.g., creativity, technical accuracy, formal tone) while maintaining general competency.
+
+## **Comparison with Existing Methods**
+
+| **Method**                   | **Features** |
+|-----------------------------|-------------|
+| **Traditional Fine-Tuning** | Updates the entire model’s weights, requiring significant computational resources and extensive training data. Enables learning new knowledge and patterns. |
+| **Lightweight Fine-Tuning (LoRA, etc.)** | Adds adaptive low-rank matrices to optimize fine-tuning. More efficient but still requires training. |
+| **Directional Enhancement (this method)** | Selectively **amplifies** the model’s intrinsic capabilities by strengthening specialized output directions. Does not introduce new knowledge. |
+
+## **Implementation Details**
+
+### **Data Preparation**
+
+Two types of datasets are used to define the specialization direction:
+- **Specialized Dataset (`enhance.txt`)**: Contains prompts focused on the capability to be enhanced.
+- **General Dataset (`normal.txt`)**: Contains diverse, neutral prompts to serve as a baseline.
+
+The difference in activations between these two datasets defines the specialization direction, ensuring that the enhancement is aligned with the target capability while preserving the model’s general functionality.
+
+### **Key Parameters**
+
+- **`instructions`**: Number of instruction samples to process (default: 128)
+- **`layer_idx`**: Index of the model layer where specialization direction is computed (default: 60% of total layers)
+- **`enhancement_factor`**: Strength of enhancement along the specialization direction (default: 1.5)
+
+### **Core Algorithm**
+
+```python
+# Compute specialization direction
+specialization_dir = specialized_mean - general_mean
+specialization_dir = specialization_dir / specialization_dir.norm()
+
+# Core part of the weight enhancement algorithm
+projection_scalars = torch.matmul(attn_output, specialization_dir)
+projection = torch.outer(projection_scalars, specialization_dir)
+enhanced_weights = attn_output + enhancement_factor * projection
+```
+
+## **Performance and Results**
+
+### **Improvements in Creative Writing Model**
+
+Experiments with creative writing models demonstrate **significant qualitative improvements**:
+
+- **Enhanced Descriptive Ability**: More vivid and detailed descriptions with richer sensory language.
+- **Improved Character Development**: Clearer character traits and more distinct personalities.
+- **Enhanced Dialogue Generation**: More natural and engaging conversational exchanges.
+- **Stronger Story Structuring**: Improved narrative flow and coherence.
+- **Increased Emotional Depth**: Greater emotional nuance and expressiveness.
+
+## **Applications**
+
+This technique can be applied to various specialized models:
+
+- **Creative Writing Models**: Optimized for novel writing, poetry, and storytelling.
+- **Educational Content Models**: Tailored for clear, structured, and pedagogical explanations.
+- **Technical Documentation Models**: Enhanced for structured and precise documentation.
+- **Business Communication Models**: Specialized for professional and formal business writing.
+- **Medical/Scientific Models**: Improved for detailed and accurate scientific explanations.
+
+## **Limitations and Future Improvements**
+
+### **Current Limitations**
+
+- **Interpretability of Specialization Directions**: Difficult to precisely determine what specific abilities are being enhanced.
+- **Single-Direction Specialization**: Currently enhances only one specific capability at a time.
+- **Control Over Enhancement Level**: The optimal enhancement factor is determined empirically.
+- **No New Knowledge Acquisition**: Cannot introduce entirely new knowledge beyond what the model already possesses.
+- **Dependence on Existing Abilities**: If the model lacks fundamental knowledge in a domain, the enhancement effects are limited.
+
+### **Future Directions**
+
+- **Multi-Directional Enhancement**: Developing techniques to enhance multiple capabilities simultaneously.
+- **Automatic Tuning**: Implementing an automated method for optimal enhancement factor selection.
+- **Interpretability of Specialization**: Researching better semantic analysis of specialization directions.
+- **User-Personalized Specialization**: Customizing specialization directions based on user preferences.
+- **Hybrid Approach**: Combining **directional enhancement** with lightweight fine-tuning to enable both ability enhancement and new knowledge learning.
+
+## **Conclusion**
+
+The **Directional Enhancement** technique provides an efficient way to strengthen specific capabilities of language models **without requiring full retraining or additional training data**. While it does not introduce new knowledge, it **amplifies latent abilities** with minimal computational cost. This method offers a practical approach for developing AI models tailored to specialized domains.
+
+## **References & Citations**
+
+This methodology was inspired by the following studies:
+
+- **Representation Engineering: A Top-Down Approach to AI Alignment**
+- **The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets**
+- **Finding Directions in GAN's Latent Space**