| # ScriptSwiftAI (Dikshan) | |
| > A QLoRA fine-tuned code generation model built and owned by **Dikshan** (India). | |
| > Specializes in generating clean, production-ready code across multiple languages and frameworks. | |
| --- | |
| ## License | |
| This model is licensed under **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)**. | |
| - ✅ You may use, modify, and distribute this model for **personal and research purposes** | |
| - ✅ You must **credit Dikshan / ScriptSwiftAI** with a link to this repository | |
| - ✅ You must **state any changes** you made | |
| - ❌ You may **NOT** use this model or any derivative for **commercial purposes** | |
| - ❌ You may **NOT** sell, sublicense, or monetize any product built on this model | |
| Full license: https://creativecommons.org/licenses/by-nc/4.0/ | |
| **Copyright © 2026 Dikshan. All rights reserved.** | |
| --- | |
| ## Model Details | |
| | Property | Value | | |
| |---|---| | |
| | **Model Name** | ScriptSwiftAI (Dikshan) | | |
| | **Base Model** | Qwen2.5-Coder-7B-Instruct | | |
| | **Fine-tuning Method** | QLoRA (4-bit quantization) | | |
| | **Training Hardware** | NVIDIA RTX 5070 Ti | | |
| | **Framework** | Hugging Face Transformers + PEFT | | |
| | **Target Steps** | ~37,500 (~2.21 epochs) | | |
| | **Dataset Size** | ~271,689 samples (11 datasets + 1,900 identity QA pairs) | | |
| | **Developer** | Dikshan (India) | | |
| | **Training Started** | June 13, 2026 | | |
| --- | |
| ## What is ScriptSwiftAI? | |
| ScriptSwiftAI is a fine-tuned AI model trained from scratch by Dikshan to generate fast, accurate, and production-ready code. It is trained on a custom curated dataset covering: | |
| - Python, JavaScript, TypeScript, React, Node.js | |
| - API development, full-stack web apps | |
| - Code explanation, debugging, and refactoring | |
| - Custom identity: responds as "ScriptSwiftAI" / "Dikshan" (not as base model) | |
| --- | |
| ## Training Configuration | |
| - **LoRA Rank:** 64, Alpha: 128 | |
| - **Quantization:** 4-bit (NF4) via bitsandbytes | |
| - **Optimizer:** paged_adamw_8bit | |
| - **Batch Size:** 2 (gradient accumulation: 8, effective batch: 16) | |
| - **Learning Rate:** 2e-4 with cosine scheduler | |
| - **Max Sequence Length:** 2048 tokens | |
| - **Warmup Steps:** 100 | |
| - **Checkpointing:** Every 500 steps with UPS auto-resume support | |
| --- | |
| ## Changelog / Training Timeline | |
| This section tracks every major checkpoint upload as proof of development history. | |
| All uploads are additionally timestamped via HuggingFace commit history. | |
| --- | |
| ### 🟢 v0.1 — Training Kickoff | |
| **Date:** June 13, 2026 | |
| **Status:** Training started | |
| - Base model loaded: Qwen2.5-Coder-7B-Instruct (Apache 2.0) | |
| - QLoRA configuration finalized | |
| - Dataset pipeline built: ~271,689 samples across 11 datasets | |
| - Identity QA pairs injected: 1,900 samples (ScriptSwiftAI / Dikshan persona) | |
| - Training infrastructure set up: train_v11_monitor.py with live loss monitoring UI | |
| --- | |
| ### 🟡 v0.2 — Best Loss Checkpoint | |
| **Date:** June 17, 2026 | |
| **Checkpoint:** checkpoint-12000 | |
| **Step:** ~12,000 / 37,500 | |
| **Loss:** 0.1920 (best recorded) | |
| **Status:** Training ongoing | |
| - Best performing checkpoint saved | |
| - Model showing strong code generation capability | |
| - Identity responses stable (responds as ScriptSwiftAI/Dikshan correctly) | |
| - Repetition penalty and top_p inference parameters tuned | |
| --- | |
| ### 🔵 v0.3 — ~77% Training Complete *(Latest)* | |
| **Date:** June 20, 2026 | |
| **Checkpoint:** checkpoint-29000 *(update with actual checkpoint number)* | |
| **Step:** ~29000+ / 37,500 | |
| **Loss:** ~0.27–0.32 (normal QLoRA batch variance) | |
| **Status:** Training ongoing | |
| - First public upload to HuggingFace | |
| - README and LICENSE added | |
| - Commercial product stack in development (Desktop + Web app) | |
| --- | |
| *This changelog will be updated with every checkpoint upload going forward.* | |
| --- | |
| ## Intended Use | |
| | Use Case | Allowed? | | |
| |---|---| | |
| | Personal coding assistant | ✅ Yes | | |
| | Research and academic use | ✅ Yes | | |
| | Open source projects | ✅ Yes (with credit) | | |
| | Commercial SaaS / products | ❌ No | | |
| | Reselling or sublicensing | ❌ No | | |
| --- | |
| ## Credit / Attribution | |
| If you use ScriptSwiftAI in any project, you must include: | |
| > *"Powered by ScriptSwiftAI, developed by Dikshan — https://huggingface.co/Dikshan1234/ScriptSwiftAI"* | |
| --- | |
| ## Contact | |
| For licensing inquiries or commercial use permissions, contact the developer directly via HuggingFace. | |
| --- | |
| *ScriptSwiftAI is an independent AI project developed entirely by Dikshan (India). | |
| --- | |
| ### 🔵 Checkpoint — Step 29,513 | |
| **Date:** June 20, 2026 at 23:50 IST | |
| **Checkpoint:** checkpoint-29513 | |
| **Step:** 29,513 / 37,500 (78.7% complete) | |
| **Loss:** 0.2475 | |
| **Epoch:** 1.738 | |
| **Zone:** 🎯 Sweet Spot | |
| **Status:** Training ongoing — auto-uploaded via train_v11_monitor.py | |