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
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
