Snapgate Surge V4
Official AI Assistant of Snapgate AI
Snapgate Surge V4 is a LoRA adapter built on top of Qwen3-14B (Dense) and instruction-tuned using proprietary conversational datasets developed by Snapgate AI.
✨ Highlights
🚀 Built on Qwen3-14B Dense
⚡ Fine-tuned using LoRA + Unsloth
💻 Optimized for coding assistance
🇮🇩 Native support for Indonesian & English
🧠 Strong instruction-following capability
📉 Final training loss: 0.0124
🤖 Official AI Assistant of Snapgate AI
🔓 Apache 2.0 Licensed LoRA Adapter
📖 Model Overview
Property Value Model Name Snapgate Surge V4 Model Type LoRA Adapter Base Model Qwen3-14B Dense Architecture Transformer Fine-tuning LoRA Framework Unsloth Trainer TRL SFTTrainer Languages English, Indonesian Context Length 2048 Tokens License Apache-2.0
🏢 About Snapgate AI
Snapgate AI is an AI platform focused on building intelligent assistants for programming, productivity, research, and everyday tasks.
The Snapgate ecosystem includes multiple AI products:
🤖 Snapgate Lite
🚀 Snapgate Titan
🧠 Cognira
📄 Ask PDF
💻 Online Code Editor
📝 AI Report Summarizer
🩺 Lexar AI
🎨 Frontend Generator
Users receive 50,000 free tokens every day .
🌐 Website
📚 Snapgate Model Family
Model Size Snapgate 3B Snapgate Code 4B Snapgate Surge V1 7B Snapgate Surge V2 10M Snapgate Surge V3 7B Active (MoE) Snapgate Surge V4 14B Dense
⚙️ Model Specifications
Specification Value Base Model Qwen3-14B Dense Fine-tuning LoRA Rank (r) 16 Alpha 32 Quantization 4-bit Framework Unsloth Trainer TRL SFTTrainer
Target Modules
q_proj
k_proj
v_proj
o_proj
gate_proj
up_proj
down_proj
✨ Capabilities
✔ Programming Assistance
Python
PHP
JavaScript
TypeScript
C++
Java
HTML/CSS
SQL
✔ Web Development
✔ Debugging
✔ Technical Writing
✔ General Question Answering
✔ Instruction Following
✔ Bilingual Conversations
✔ Long-form Explanations
📈 Training Progress
The model was instruction-tuned using Unsloth and TRL SFTTrainer .
Training Summary
Metric Value Framework Unsloth Trainer TRL SFTTrainer Total Steps 300 Completed 300 / 300 (100%) Epochs 3 Initial Loss 1.7304 Final Loss 0.0124
Training Loss
Step Loss 10 1.7304 20 0.4045 30 0.2602 40 0.1483 50 0.0915 60 0.0517 70 0.0316 80 0.0266 90 0.0250 100 0.0175 150 0.0152 200 0.0141 250 0.0135 300 0.0124
Training Notes
✅ Successfully completed all scheduled training steps.
📉 Stable convergence throughout training.
💾 Periodic checkpoints were saved to support training resumption.
⚡ Optimized using Unsloth for efficient fine-tuning.
📦 Repository Contents
This repository contains:
✅ LoRA adapter
✅ Adapter configuration
✅ Tokenizer configuration
✅ Generation configuration
✅ Model Card
This repository DOES NOT contain the Qwen3-14B base model.
📋 Requirements
Install dependencies:
pip install unsloth transformers peft accelerate bitsandbytesRecommended versions:
transformers >= 4.52
peft >= 0.16
torch >= 2.6
unsloth
🚀 Quick Start
1 from unsloth import FastLanguageModel
2
3 model , tokenizer = FastLanguageModel . from_pretrained (
4 model_name = "kadalicious22/snapgate-surge-v4" ,
5 max_seq_length = 2048 ,
6 load_in_4bit = True ,
7 )
8
9 FastLanguageModel . for_inference ( model )
10
11 messages = [
12 {
13 "role" : "system" ,
14 "content" : "You are Surge, an AI assistant developed by Snapgate AI."
15 } ,
16 {
17 "role" : "user" ,
18 "content" : "Who are you?"
19 }
20 ]
21
22 inputs = tokenizer . apply_chat_template (
23 messages ,
24 tokenize = True ,
25 add_generation_prompt = True ,
26 return_tensors = "pt"
27 ) . to ( "cuda" )
28
29 outputs = model . generate (
30 input_ids = inputs ,
31 max_new_tokens = 256 ,
32 temperature = 0.7 ,
33 top_p = 0.9 ,
34 )
35
36 print ( tokenizer . decode ( outputs [ 0 ] [ inputs . shape [ 1 ] : ] , skip_special_tokens = True ) )
🔧 Loading the LoRA Adapter
This repository contains only the LoRA adapter .
Download the Qwen3-14B base model and load the adapter:
1 from unsloth import FastLanguageModel
2
3 model , tokenizer = FastLanguageModel . from_pretrained (
4 model_name = "unsloth/Qwen3-14B" ,
5 max_seq_length = 2048 ,
6 load_in_4bit = True ,
7 )
8
9 model . load_adapter ( "kadalicious22/snapgate-surge-v4" )
10
11 FastLanguageModel . for_inference ( model )
🔀 Merge the Adapter
To create a standalone model:
1 from unsloth import FastLanguageModel
2
3 model , tokenizer = FastLanguageModel . from_pretrained (
4 model_name = "unsloth/Qwen3-14B" ,
5 load_in_4bit = False ,
6 )
7
8 model . load_adapter ( "kadalicious22/snapgate-surge-v4" )
9
10 model = model . merge_and_unload ( )
11
12 model . save_pretrained_merged (
13 "snapgate-surge-v4-merged" ,
14 tokenizer ,
15 save_method = "merged_16bit" ,
16 )
Load the merged model normally:
1 from transformers import AutoModelForCausalLM
2 from transformers import AutoTokenizer
3
4 tokenizer = AutoTokenizer . from_pretrained (
5 "snapgate-surge-v4-merged"
6 )
7
8 model = AutoModelForCausalLM . from_pretrained (
9 "snapgate-surge-v4-merged" ,
10 torch_dtype = "auto" ,
11 device_map = "auto" ,
12 )
🌐 Snapgate API
1 curl https://snapgate.tech/api/v1/chat/completions \
2 -H "Content-Type: application/json" \
3 -H "Authorization: Bearer YOUR_API_KEY" \
4 -d '{
5 "model":"snapgate-lite",
6 "messages":[
7 {
8 "role":"user",
9 "content":"Hello!"
10 }
11 ]
12 }'
📊 Training Details
Base Model: Qwen3-14B Dense
Fine-tuning: LoRA
Framework: Unsloth
Trainer: TRL SFTTrainer
Dataset: Proprietary Snapgate AI conversational dataset
Languages: English & Indonesian
Focus:
Coding
Instruction Following
Technical Assistance
General Conversations
⚠️ Notes
This repository only contains the LoRA adapter.
To use the model, you must first obtain the original Qwen3-14B base model and then load this adapter using PEFT or Unsloth .
📄 License
Released under the Apache 2.0 License.
The Qwen3-14B base model remains subject to its original license.
📚 Citation
1 @misc{snapgate2026surgev4,
2 title={Snapgate Surge V4},
3 author={Snapgate AI},
4 year={2026},
5 publisher={Hugging Face},
6 howpublished={\url{https://huggingface.co/kadalicious22/snapgate-surge-v4}}
7 }
🔗 Links
Website
About
Hugging Face
Instagram
Discord
Base Model
⭐ If you find Snapgate Surge V4 useful, consider giving the repository a Star ❤️ on Hugging Face to support future development.
