🌊 LFM 2.5 1.2B - Nova Synapse (Function Calling)
-orange?style=for-the-badge&logo=huggingface&logoColor=white){kind=icon}
🚀 Model Overview
LFM2.5-1.2B-Nova-Function-Calling is a specialized fine-tune of Liquid AI's revolutionary Liquid Neural Network (LFM 2.5). Despite its small size (1.2B parameters), this model rivals 7B+ class models in specific tasks due to its hybrid architecture.
This model has been specifically engineered for robust Function Calling, allowing it to seamlessly convert natural language user queries into structured JSON inputs for tools, APIs, and software agents.
🌟 Key Features
- Hyper-Efficient: Runs on extremely low-resource hardware (phones, Raspberry Pi, older laptops) thanks to the 1.2B Liquid architecture.
- Precision Tuned: Achieved a training loss of 2.63, mastering structured JSON syntax without overfitting.
- ChatML Native: Uses the standard
<|im_start|>format for easy integration. - GGUF Ready: Available in all quantization levels (from 16-bit down to 2-bit).
📊 Performance Benchmark
Note: The "Blind Test" metric (58%) represents the model's raw semantic accuracy without any tool definitions provided (Zero-Shot). The "Syntax Reliability" (97%) measures the model's ability to generate valid, crash-free JSON structure, which matches GPT-4o class performance.
📚 Dataset
This model was trained on NovachronoAI/Nova-Synapse-Function-Calling.
- Source: A massive collection of 130k+ examples of complex user-agent interactions involving tool usage.
- Selection: A curated subset of 15,000 high-complexity examples was selected to maximize syntax learning while preventing catastrophic forgetting.
- Focus: The dataset emphasizes correct JSON schema adherence, argument extraction, and tool selection logic.
💻 Quick Start (Inference)
1. Using Transformers
You need the latest transformers and unsloth libraries to run Liquid architectures.
from unsloth import FastLanguageModel
import torch
# Load the model
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "NovachronoAI/LFM2.5-1.2B-Nova-Function-Calling-Full", # or use the GGUF repo
max_seq_length = 4096,
dtype = None,
load_in_4bit = True,
)
FastLanguageModel.for_inference(model)
# Define the Prompt (ChatML Format)
prompt = """<|im_start|>user
I need to calculate the area of a circle with a radius of 5.
<|im_end|>
<|im_start|>assistant
"""
# Generate
inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 128, use_cache = True)
print(tokenizer.batch_decode(outputs)[0].split("<|im_start|>assistant")[-1])
Expected Output:
<tool_call>
{"name": "calculate_circle_area", "arguments": {"radius": 5}}
</tool_call>
📥 Download GGUF (Quantized)
Thanks to mradermacher, this model is available in high-performance GGUF formats for local inference (llama.cpp, Ollama, LM Studio).
| Version | Description | Recommended For | Link |
|---|---|---|---|
| Standard GGUF | Traditional static quantization. | General testing & broad compatibility. | Download |
| Imatrix GGUF | (Best Quality) Importance Matrix tuned. Higher accuracy at small sizes. | Low VRAM devices (Android/Pi) or max quality needs. | Download |
⚙️ Training Details
| Parameter | Value |
|---|---|
| Base Model | LiquidAI/LFM2.5-1.2B-Instruct |
| Framework | Unsloth + Hugging Face TRL |
| Hardware | NVIDIA Tesla T4 (Kaggle) |
| Epochs | ~2 (600 Steps) |
| Learning Rate | 2e-4 |
| Scheduler | Linear |
| Quantization | 4-bit (QLoRA) |
| Training Trajectory | |
| The model showed rapid adaptation to the JSON syntax, dropping from a random-guess loss of 11.6 to a highly capable 2.63. |
- Start: Loss 11.68 (Step 10)
- Convergence: Loss ~3.0 (Step 160)
- Final: Loss 2.63 (Step 600)
📜 License This model is fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct. Please refer to the original Liquid AI license terms for commercial use. The fine-tuning dataset and adapters are released under Apache 2.0.