|
|
--- |
|
|
license: other |
|
|
license_name: lfm1.0 |
|
|
license_link: https://huggingface.co/LiquidAI/LFM2-2.6B/blob/main/LICENSE |
|
|
metrics: |
|
|
- magic judge |
|
|
base_model: |
|
|
- LiquidAI/LFM2-2.6B |
|
|
tags: |
|
|
- lmstudio |
|
|
- madlabOSS |
|
|
- magic judge |
|
|
--- |
|
|
|
|
|
# LMS Guide 2.6b |
|
|
|
|
|
## 🧠 Overview |
|
|
The **LMS Guide 2.6b** is part of the **MadlabOSS LM Studio Guide** family — a lineup of small, efficient, and highly aligned assistant models trained specifically to provide deterministic, hallucination‑resistant guidance for LM Studio users. |
|
|
|
|
|
This model is trained on a curated dataset of LM Studio–specific instructions, workflows, troubleshooting steps, and conceptual explanations. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 Intended Use |
|
|
This model is optimized for: |
|
|
|
|
|
- LM Studio onboarding |
|
|
- workflow explanations |
|
|
- feature descriptions |
|
|
- troubleshooting guidance |
|
|
- plugin/server integration help |
|
|
- safe, deterministic assistant behavior |
|
|
|
|
|
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: |
|
|
|
|
|
- **6,000+ LM Studio–specific instruction/response pairs** |
|
|
- Clean, domain‑specific, ontology‑consistent data |
|
|
- Minor general‑purpose conversational data |
|
|
- No web‑scraped content |
|
|
- Full LM Studio Documentation |
|
|
|
|
|
A 36k+ expanded dataset is planned for v2.0. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🏋️ Training Procedure |
|
|
|
|
|
### **Hyperparameters** |
|
|
- Epochs: 6 |
|
|
- Batch size: 16 |
|
|
- Learning rate: cosine schedule, peak ~4e‑5 |
|
|
- Optimizer: AdamW |
|
|
- Gradient clipping: 1.0 |
|
|
- Gradient accumulation: 1 |
|
|
|
|
|
### **Hardware** |
|
|
Training was performed on: |
|
|
|
|
|
- RTX 6000 Ada (96GB) (1.2b + 2.6b) |
|
|
- Dual RTX 3090 (Magic Judge) |
|
|
- RTX 3070 (for 0.35B + 0.7b) |
|
|
|
|
|
--- |
|
|
|
|
|
## 📊 Evaluation |
|
|
|
|
|
### **Judge Score** |
|
|
Semantic correctness, ontology adherence, and hallucination resistance. |
|
|
|
|
|
|
|
|
### **Qualitative Behavior** |
|
|
- Strong adherence to LM Studio terminology |
|
|
- Low hallucination rate |
|
|
- Deterministic, predictable responses |
|
|
- Not optimized for open‑domain reasoning |
|
|
|
|
|
--- |
|
|
|
|
|
## 🔒 Safety |
|
|
This model is trained exclusively on LM Studio–specific content. |
|
|
It avoids hallucinating non‑existent LM Studio features and adheres to a strict ontology. |
|
|
|
|
|
It is **not** designed for: |
|
|
|
|
|
- political content |
|
|
- 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 LM Studio scope |
|
|
- Static accuracy metrics underestimate real performance |
|
|
|
|
|
--- |
