---
license: other
license_name: lfm1.0
license_link: https://huggingface.co/LiquidAI/LFM2-2.6B/blob/main/LICENSE
metrics:
- Synthetic Data Expansion Benchmark
base_model:
- LiquidAI/LFM2-2.6B
tags:
- lmstudio
- madlabOSS
- synthetic data generator
---

# Madlab Synthetic Data Generator

## 🧠 Overview
The **Madlab SDG 2.6B** is part of the **MadlabOSS Synthetic Data Generator** family — a suite of small, efficient synthetic data generators designed for rule‑consistent, semantically coherent variation.  
This model was trained on a closed-source dataset created through a multi-stage synthetic data generation process using a modified Madlab training pipeline.

---

## 🚀 Intended Use
This model is optimized for:

- Madlab synthetic data generation

It is **not** intended as a general-purpose chatbot.

---

## 🧩 Model Details

**Base Model:** LFM2-2.6B  
**Parameter Count:** 2.6 Billion  
**Training Type:** Supervised fine-tuning  
**Sequence Length:** 1024  
**Precision:** FP16  
**Framework:** PyTorch / Transformers  

---

## 📦 Training Data
The model was trained on:

- **1444 compressed and encoded dataset pairs**  
- High variation in output  
- Preservation of semantic meaning  
- Data entirely generated with Madlab  

---

## 🏋️ Training Procedure

### **Hyperparameters**
- Epochs: 6  
- Batch size: 48  
- Learning rate: cosine schedule, peak ~4e-5  
- Optimizer: AdamW  
- Gradient clipping: 1.0  
- Gradient accumulation: 1  

### **Hardware**
Training was performed on:

- RTX 6000 Blackwell (96GB)

---

## 📊 Evaluation

### **Synthetic Data Expansion Benchmark**
A curated set of 30 input/target pairs was programmatically expanded using a Python script.  
Metrics include seed pairs covered, total variation count, and semantic quality.  
The task is to generate 5 variations of each incoming pair.

| Run | Model | Semantic Quality | Variations | Seeds Covered | Efficiency (Variations/Param) | Dataset |
|-----|------------|---------------|------------|---------------|-------------------------------|--------------|
| 1 | LFM2-350M-16 | 6.5 | 94 | 23 | 268.57 | Madlab sdg small |
| 2 | LFM2-350M-16 | 3.5 | 46 | 11 | 131.43 | base model |
| 3 | LFM2-350M-f16 | 6.5 | 97 | 22 | 277.14 | Madlab sdg small |
| 4 | Qwen3-coder-30B-instruct-q8 | 8.2 | 149 | 26 | 4.97 | base model |
| 5 | LFM2-350M-f16 | 7.5 | 136 | 21 | 388.57 | Madlab sdg medium |
| 6 | LFM2-2.6B-f16 | 9.0 | 137 | 25 | 52.69 | Madlab sdg medium |
| 7 | LFM2-2.6B-f16 | 9.9 | 180 | 25 | 69.23 | Madlab sdg large |
| 8 | LFM2-2.6B-f16 | 6.2 | 157 | 20 | 60.38 | Madlab sdg test |
| 9 | LFM2-2.6B-f16 | 10.0 | 248 | 27 | 95.38 | Madlab sdg large |
| 10 | Qwen3-235B-q3-k_m | 9.5 | 150 | 27 | 0.64 | base model |
| 11 | LFM2.5-1.2B-instruct-f16 | 9.1 | 244 | 30 | 203.33 | Madlab sdg large |

### **Qualitative Behavior**
- Overperforms in variation count  
- Maintains strict semantic correctness  

---

## 🔒 Safety
This model is a synthetic data generator. It is not designed for conversational use and is not suitable for anything other than generating synthetic datasets.

It is **not** designed for:

- Political advice  
- Medical advice  
- Legal advice  
- General-purpose conversation  

---

## ⚠️ Limitations
- Not a general assistant  
- Not trained for coding, math, or open-domain reasoning  
- May refuse tasks outside the Madlab SDG scope  

---