Tiranyx/sidix-lora

SIDIX LoRA — archived research build (April–August 2026)

Status: archived, not maintained. SIDIX was a self-hosted AI agent project. Its research build ended in August 2026, and the hosted app was retired on purpose. Its research continued as MiganCore.

What it is

• Base: Qwen/Qwen2.5-7B-Instruct (Apache-2.0).
• Method: QLoRA, r = 16, α = 32, on all seven projection modules. Trained on one Kaggle T4. (An earlier version of this card listed r = 64 on four modules; that was wrong.)

Training data (713 rows, 641 used for training)

• 673 auto-generated question–answer pairs. Template questions are answered with verbatim excerpts from the project's corpus:
  • research notes, drafted with the help of AI assistants;
  • principles, sources and bibliography;
  • 99 excerpts clipped from third-party web pages, which may carry their own licence obligations, such as CC BY-SA.
• 40 hand-written pairs.

Because training data came from the outputs of third-party commercial LLM services, check the terms of those services before any use beyond research.

Evaluation

No pre-registered evaluation was published for this adapter.

Code

• The full source is at commit c1185b7 and tag v0.7.0 of fahmiwol/sidix. The main branch now holds the archive README and the Quran Lab provider.
• Successor research: fahmiwol/migancore.

License: MIT for the adapter; the base model is Apache-2.0. Author: Fahmi Ghani (fahmiwol@gmail.com).