GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jhalakdev67dev
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
jhalakdev67dev
Model Tree
sarvamai/sarvam-1
Input Modalities
Output Modalities
Supported Functionality
A LoRA adapter for Sarvam-2B (India's open-source LLM, Apache 2.0) that turns it into AURA — a system agent that understands Hindi, Hinglish and English commands and executes real actions on the machine:
| Skill | What it does |
|---|---|
| 📁 Aura Manager | create / save / delete / move / organize files |
| 📦 Aura Installer | install & uninstall software (guarded) |
| 🚀 Aura Launchpad | open apps, websites, browser searches |
| 🔍 Aura Researcher | web search & page fetch |
| 🧮 Aura Aryabhatta | math via a Python sandbox |
| ⚙️ Aura Controller | screenshots, clipboard, volume, system info |
Plus personality: AURA replies in natural Hinglish, and knows when to just talk instead of calling a skill.
rm -rf /, format, fork bombs...) are hard-blocked| Method | QLoRA (r=16, alpha=32) + NEFTune (alpha=5) + prompt-masked loss |
| Data | 6,687 examples — 6 skills + chitchat, 3 languages (Hindi / Hinglish / English), incl. multi-turn [TOOL RESULT] → confirmation patterns |
| Hardware | Free Kaggle T4 x2 — total GPU budget: ₹0 |
| Loss | 1.77 → 0.0072 in 160 steps (converged; a 250x drop) |
This checkpoint: step-164 draft of the first run. The official full 1-epoch run + router-accuracy eval is on the way and will replace this.
The full agent runtime (skills, guardrails, execution loop, self-healing) lives in the open-source one-file repo — download AURA_onefile.py and run:
bash
python AURA_onefile.py agent --model sarvamai/sarvam-2b --adapter ./aura_loraTo load the adapter yourself:
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("sarvamai/sarvam-2b", device_map="auto")model = PeftModel.from_pretrained(base, "jhalakdev67dev/aura-v1-lora")tok = AutoTokenizer.from_pretrained("jhalakdev67dev/aura-v1-lora")Note: the adapter expects the AURA system prompt (see
build_system_prompt()in the one-file repo) — plain chat without it will still work but the skill-routing is trained for the agent format.
Trained by Jhalak Jhajhria (13, India) — built in public on free compute.
Base model: Sarvam-2B by Sarvam AI (Apache 2.0) · Trained with PEFT + Transformers · Compute: Kaggle T4
AURA — flow. resonance. connection. 👻
python AURA_onefile.py agent --model sarvamai/sarvam-2b --adapter ./aura_lorafrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("sarvamai/sarvam-2b", device_map="auto")model = PeftModel.from_pretrained(base, "jhalakdev67dev/aura-v1-lora")tok = AutoTokenizer.from_pretrained("jhalakdev67dev/aura-v1-lora")