| --- |
| license: apache-2.0 |
| tags: |
| - text-generation-inference |
| - transformers |
| - code |
| - agent |
| - text-generation |
| - custom-tune |
| - slerp-merge |
| - ties-merge |
| - reasoning |
| base_model: |
| - microsoft/phi-4 |
| - microsoft/Phi-4-reasoning-plus |
| language: |
| - en |
| pipeline_tag: text-generation |
| --- |
| |
| <div align="center"> |
|
|
| # 🇮🇳 Ekant-14B-small (Agentic Reasoning Edition) |
|
|
| ## 🌟 **Made in India** 🌟 |
|
|
| <h3>🚀 A High-Performance Specialist Model Fused with Deep Reasoning</h3> |
|
|
| <p align="center"> |
| Developed by <b>Jagneshdeveloper</b> |
| </p> |
|
|
| --- |
|
|
| `📄 License: Apache 2.0` | `⚙️ Parameters: 14 Billion` | `💻 Focus: Elite Coding, Reasoning & Agents` |
|
|
| </div> |
|
|
| --- |
|
|
| ## 📌 Overview |
|
|
| **Ekant-14B-small** is an advanced 14-billion parameter large language model proudly developed by **Jagneshdeveloper**. While initially initialized via custom-trained adapter matrices, this final artifact is a **fully unquantized standalone model** in true `float16` precision. |
|
|
| Built on top of the powerful **microsoft/phi-4** architecture, this model has been custom-engineered and cross-compiled across multiple advanced mathematical optimization passes (including **SLERP** and **TIES** multi-model fusion protocols) to integrate elite agentic logic with deep, multi-step validation tracking. |
|
|
| --- |
|
|
| ## 🔬 Fusing & Pipeline Lifecycle |
|
|
| This model was compiled under a strict resource-constrained hardware architecture using custom disk-free sharded watchdog pipelines to guarantee full float precision mapping without accuracy loss: |
| 1. **LoRA Fine-Tuning**: Initial specialized instruction sets were targeted and compiled into low-rank matrix sub-layers (`ekant-adapter`). |
| 2. **Base Integration**: Unquantized adapter weights were chemically baked directly back into the core 29.3GB `microsoft/phi-4` tensor layers. |
| 3. **Vanilla Alignment**: Merged via **SLERP** (Spherical Linear Interpolation) at a calibrated `0.6/0.4` ratio back with the foundational base model to act as a stabilizing anchor and counteract catastrophic forgetting. |
| 4. **Deep Reasoning Injection**: Fused via **TIES** (Trimming, Electing, and Merging) to combine our custom capabilities directly with reinforcement-learned o3-style logic pathways. |
|
|
| --- |
|
|
| ## ⚡ Key Capabilities |
|
|
| * **🧠 Deep Reasoning plus**: Features integrated reflection traces, error self-correction blocks, and highly granular problem-solving structures. |
| * **💻 Coding Specialist**: Optimized to write, debug, analyze, and safely refactor high-complexity code structures across Python, JavaScript, C++, Rust, and Go. |
| * **🤖 Agentic Excellence**: High structural accuracy for tool-use, functional api execution loops, and generating strictly formatted outputs (like complex JSON or nested system commands). |
|
|
| --- |
|
|
| ## 📊 Model Summary |
|
|
| * **Model Name:** Ekant-14B-small (Agentic Ultimate Edition) |
| * **Developer:** Jagneshdeveloper |
| * **Base Architecture:** Built on top of Microsoft Phi-4 (Phi3 For Causal LM Core Class) |
| * **Parameters:** 14 Billion (14B) |
| * **License:** Apache 2.0 (Permissive Open-Source) |
| * **Primary Language:** English (en) |
|
|
| --- |
|
|
| ## 💻 Quick Start |
|
|
| You can quickly load and deploy **Ekant-14B-small** using the Hugging Face `transformers` library: |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| # Real repository target path verified on your profile |
| model_name = "Jagneshdeveloper/ultimate-Ekant-14b" |
| |
| # Load the optimized tokenizer and model |
| tokenizer = AutoTokenizer.from_pretrained(model_name) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_name, |
| device_map="auto", |
| torch_dtype=torch.float16, |
| trust_remote_code=True |
| ) |
| |
| # Test prompt for deep reasoning & agentic execution |
| prompt = "Write an optimized Python script to scrape website data dynamically, handle API authentication token refreshes, and format it into a structured JSON array." |
| inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
| outputs = model.generate( |
| **inputs, |
| max_new_tokens=512, |
| temperature=0.5, |
| do_sample=True, |
| pad_token_id=tokenizer.eos_token_id |
| ) |
| |
| print(tokenizer.decode(outputs, skip_special_tokens=True)) |
| ``` |
|
|
| --- |
|
|
| ## 🛠️ Intended Uses & Limitations |
|
|
| ### Ideal Use Cases |
| * Building autonomous AI agents and automated API execution loops. |
| * Serving as a local or cloud-hosted programming and mathematical reasoning assistant. |
| * Handling multi-step logical text generation and complex data extraction tasks. |
|
|
| ### Limitations |
| * Standard 14B computing constraints apply; heavy inference calls may require sharding or quantization depending on available VRAM allocations. |
| * Users should verify complex logic outputs before running generated scripts straight into a live production workspace. |
|
|
| --- |
|
|
| ## 🤝 Attribution & Support |
|
|
| Created with ❤️ by **Jagneshdeveloper** in India. This model is distributed under the open and permissive **Apache 2.0 License**, providing full freedom for commercial deployment, modifications, and distributed derivatives. |
|
|
| Special credit and attribution are extended to **Microsoft** for their foundational open-weights research contributions (`phi-4` and `Phi-4-reasoning-plus`), which served as the essential structural pillars and base anchors for this advanced mathematical crossover fusion project. |
|
|
| For feedback, feature requests, or collaborations, feel free to open a discussion in the community tab! |
|
|
