SIDIX LoRA — Free & Open Source AI Agent

"Thinks, Learns & Creates."

SIDIX is an autonomous AI agent that runs 100% locally — no per-query cost, no data leaves your server, no vendor lock-in.

This repository hosts the LoRA adapter fine-tuned on top of Qwen/Qwen2.5-7B-Instruct using QLoRA (4-bit NF4 quantization, Kaggle T4 GPU). It is the language-model component of the larger SIDIX agent system, which adds RAG, 48 tools, ReAct loop, semantic cache, and a continuous-learning growth loop on top.

🔗 GitHub (full agent): github.com/fahmiwol/sidix 🌐 Live Demo (Free): app.sidixlab.com 📜 License: MIT 🏷️ Latest version: v2.1.4 (Vol 20-fu3, 2026-04-26)

What's Different about SIDIX?

Property	SIDIX	Closed-Vendor AI
Inference cost	Free (own GPU/CPU)	Per-token billing
Data egress	Stays on your server	Sent to vendor cloud
Open-source	✅ MIT — adapter + agent + corpus	Closed
Self-hostable	✅ End-to-end	❌
Vendor LLM API fallback	❌ Never (`brain_qa` core has zero `openai`/`anthropic`/`google` imports)	N/A
Epistemic labeling	`[FACT] / [OPINION] / [SPECULATION] / [UNKNOWN]` on sensitive claims	Generic disclaimer
Growth loop	LoRA retrains nightly from corpus queue	Static snapshot

Architecture (One Glance)

┌──────────────────────────────────────────────────────────┐
│  brain_qa (FastAPI, VPS or your laptop)                  │
│  ─ ReAct loop · 48 tools · sanad chain                   │
│  ─ Semantic cache (BGE-M3 embedding, per-domain TTL)     │
│  ─ Complexity router (simple / standard / deep)          │
│  ─ Domain detector (fiqh / medis / coding / factual)     │
└──────────────────────────────────────────────────────────┘
                          ↓ HTTP
┌──────────────────────────────────────────────────────────┐
│  vLLM (this LoRA on Qwen2.5-7B-Instruct base)            │
│  ─ Self-hosted (RunPod serverless / your GPU)            │
│  ─ Returns generative output, ReAct + RAG handled by     │
│    brain_qa above                                         │
└──────────────────────────────────────────────────────────┘

The adapter alone gives you SIDIX's voice and behavior. For the full agent (RAG, tools, growth loop) clone the GitHub repo — that ties this adapter into the orchestration layer.

Quick Start

A. Standalone (transformers + peft)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
)
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    quantization_config=bnb,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "Tiranyx/sidix-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

# SIDIX uses Qwen2.5 chat template + persona-flavored system prompt
messages = [
    {"role": "system", "content":
        "Kamu adalah SIDIX, AI agent yang jujur dan bersumber. "
        "Persona aktif: AYMAN (general/hangat). Jawab natural; untuk klaim "
        "fiqh/medis/data, beri label [FAKTA]/[OPINI]/[TIDAK TAHU] + sumber."},
    {"role": "user", "content": "Apa beda RAG sama fine-tuning untuk knowledge update?"},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok([text], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

B. Via vLLM (Production)

vllm serve Qwen/Qwen2.5-7B-Instruct \
  --enable-lora \
  --lora-modules sidix=Tiranyx/sidix-lora \
  --max-lora-rank 64

C. Full agent (RAG + tools + growth loop)

git clone https://github.com/fahmiwol/sidix && cd sidix
# Follow README — quick start ~5 min

Training Details

Parameter	Value
Base model	`Qwen/Qwen2.5-7B-Instruct`
Method	QLoRA (4-bit NF4)
LoRA rank	64
LoRA alpha	128
Target modules	`q_proj`, `v_proj`, `k_proj`, `o_proj`
Training GPU	Kaggle T4 (15 GB)
Corpus	SIDIX research notes ~1,182 docs (curated)
Domains	Islamic epistemology, agentic AI research, creative copy, coding, brand strategy, content planning
Languages	Indonesian (primary), English, Arabic
Cadence	Nightly retrain from filtered corpus queue (BadStyle filter, see Vol 20-fu2)

5 Personas (LOCKED 2026-04-26)

SIDIX adapts voice, depth, and framing based on which persona is active. Each has a distinct way of speaking — not boilerplate variations of one prompt.

Persona	Specialty	Pronoun cue	Example query
UTZ	Creative, visual, brainstorm	"aku"	"Bantuin desain logo coffee shop yang feel warm + handcrafted"
ABOO	Engineer, code review, debug	"gue"	"Audit fungsi Python ini, ada race condition?"
OOMAR	Strategist, business, decision	"saya"	"Bandingkan strategi pricing freemium vs flat untuk SaaS B2B"
ALEY	Researcher, fiqh, deep academic	"saya"	"Hukum puasa di hari Senin Kamis menurut 4 mazhab?"
AYMAN	General, warm chat, everyday	"halo"	"Lagi galau soal pilihan karir, bisa ngobrol bareng?"

The agent layer (in fahmiwol/sidix) auto-routes based on question signal, or honors manual selection.

48 Tools (Agent Layer)

The LoRA alone is a chat model. The agent layer wraps it with 48 active tools:

Knowledge:  search_corpus · read_chunk · list_sources · concept_graph
Web:        web_fetch · web_search · pdf_extract
Code:       code_sandbox · code_analyze · code_validate · project_map
Creative:   generate_copy · brand_kit · plan_campaign · generate_ads
Image:      text_to_image (SDXL self-hosted)
Meta:       self_inspect · orchestration_plan · muhasabah_refine · tadabbur_*
Growth:     prompt_optimizer · roadmap_* · workspace_* · learn_agent
Admin:      complexity_tier · domain_detect · semantic_cache_stats

ReAct picks the right tool per turn — see apps/brain_qa/brain_qa/agent_react.py.

What's New (Vol 14 → Vol 20)

Vol	Feature	Notes
20-fu3	Simple-tier fast-path	Greetings/ack `78s → 2s` (37× speedup)
20	Semantic cache (L1 + L2) + BGE-M3 embedding	Multilingual ID, per-domain TTL
20	Complexity router (`simple/standard/deep`)	Auto-route reasoning depth
20	Domain detector (fiqh/medis/coding/factual)	Per-domain cache threshold + sanad gating
20	Style anomaly filter (BadStyle defense)	Corpus poisoning prevention
19	Relevance + Quality Sprint	Retrieval ranking improvements
17	CodeAct enrich + MCP wrap	Code blocks auto-execute
16	Creative Agent Ecosystem	10 domain × 37 agent (debate/iteration)
15	LoRA SIDIX adapter (this repo)	First public release on Qwen2.5-7B

Full version history: CHANGELOG.md.

Privacy & Security

✅ Zero data egress to external servers — all inference local
✅ No vendor LLM API key required — brain_qa core has zero openai/anthropic/google imports
✅ 4-label epistemic tagging — hallucinations get labeled, not hidden
✅ Identity masking — backbone provider names never confirmed/denied in user-facing output
✅ MIT License — free to use, modify, and redistribute

Limitations & Honest Notes

Indonesian-first: Best quality in Indonesian. English second, Arabic third.
Not a benchmark hero: Optimized for epistemic honesty and agent loops, not MMLU/HumanEval leaderboards. We chose calibration over scores.
Sanad chains depend on corpus: For fiqh/medis claims, quality of citation depends on the corpus you load. The default corpus emphasizes Islamic epistemology and AI research.
Adapter alone ≠ agent: This repo is the LoRA. The 48-tool ReAct loop, RAG, growth loop, and semantic cache live in fahmiwol/sidix. Use them together for the full SIDIX experience.

Contribute

GitHub PR: add a research note to brain/public/research_notes/
Live feedback: try it on app.sidixlab.com, tag what's wrong
Fork: github.com/fahmiwol/sidix

Citation

@software{sidix2026,
  title={SIDIX: Free \& Open Source Autonomous AI Agent},
  author={SIDIX Project},
  year={2026},
  url={https://github.com/fahmiwol/sidix},
  license={MIT}
}

sidixlab.com · "We don't build AI that replaces human judgment. We build AI that makes human judgment more informed."

Downloads last month: 71

Model tree for Tiranyx/sidix-lora

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Adapter

(2035)

this model