---
license: apache-2.0
base_model: unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit
tags:
- agentic
- function-calling
- tool-use
- tool-router
- low-latency
- unsloth
- tensor-type:fp16
metrics:
- accuracy
pipeline_tag: text-generation
---

# Aether-1.5B-Agentic-core

<p align="left">
  <img src="https://img.shields.io/badge/IDENTITY-AETHER--1.5B--AGENTIC--CORE-cyan?style=for-the-badge&logo=cpu" alt="Identity: AETHER-1.5B-AGENTIC-CORE">
</p>

**Aether-1.5B-Agentic-core** is an elite, edge-native language model explicitly optimized to serve as the **core routing hub and tool-execution engine** for autonomous multi-agent orchestration frameworks (such as CrewAI, LangChain, and AutoGen). 

By optimizing the attention heads specifically for API schemas, this model bridges a massive architectural gap: enabling localized, low-latency deployment without sacrificing structural integrity during code output steps.

---

## ⚡ Technical Specifications

| Feature | System Configuration |
| :--- | :--- |
| **Model Blueprint** | Qwen-2.5 (Instruct variant backplane) |
| **Parameter Volume** | 1.54 Billion |
| **Context Window** | 4,096 Tokens |
| **Quantization Format** | Un-quantized; natively merged back into **16-bit Float (fp16)** |
| **Inference VRAM Profile** | ~3.5 GB (Highly accessible for consumer hardware) |
| **Primary Specialty** | Deterministic JSON-Schema Parsing & Argument Extraction |

---

## 🛠️ The Fine-Tuning Pipeline

Standard small language models (<3B parameters) are notoriously fragile when handling code parameters—they frequently introduce syntax hallucinations, drop trailing brackets, or introduce verbose conversational fluff (*"Sure, let me call that function for you!"*) that instantly breaks automated software parsers.

### 🧠 Training Strategy & Hyperparameters
**Aether-1.5B-Agentic-core** was constructed via Parameter-Efficient Fine-Tuning (PEFT) using **Unsloth**:
* **Quantization Method:** 4-bit QLoRA target module injection to maximize gradient headroom.
* **Target Modules:** Complete attention mechanism coverage (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`).
* **Dataset Alignment:** Conditioned exclusively on high-fidelity multi-turn function invocation layouts from `NousResearch/hermes-function-calling-v1`.
* **Baking Protocol:** Merged permanently back into structural base layers (`merged_16bit`), eliminating adapter latency overhead entirely.

---

## ⚙️ System Infrastructure & Stack Badges

<p align="left">
  <img src="https://img.shields.io/badge/Optimization-Unsloth-green?style=flat-square&logo=python&logoColor=white" alt="Unsloth Optimized">
  <img src="https://img.shields.io/badge/Framework-PyTorch-EE4C2C?style=flat-square&logo=pytorch&logoColor=white" alt="PyTorch Framework">
  <img src="https://img.shields.io/badge/Dataset-Hermes%20Function-blue?style=flat-square&logo=databricks&logoColor=white" alt="Hermes Dataset">
  <img src="https://img.shields.io/badge/License-Apache%202.0-red?style=flat-square&logo=apache" alt="Apache 2.0 License">
  <img src="https://img.shields.io/badge/Compute-T4%20GPU-orange?style=flat-square&logo=nvidia&logoColor=white" alt="Compute Footprint">
</p>

---

## 🚀 Behavioral Traits & Core Abilities

1. **Deterministic Structured Layouts:** Hardened against schema syntax decay, guaranteeing valid, clean, parseable JSON payload extractions.
2. **Zero-Dialogue Overhead:** Stripped of non-operational text arrays. The model targets raw arguments instantly, cutting down execution latency and compute token costs.
3. **Strict Data-Type Preservation:** Natively correlates natural text variables into explicit system-level parameters (e.g., matching raw strings directly to accurate `int`, `boolean`, or `array` properties).

---

## 💻 Implementation & Inference Sample

To trigger the tool-calling pathway natively, structure your system prompt with clear tool boundaries.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Jenil05/Aether-1.5B-Agentic-core"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

messages =