██╗ ██╗ █████╗ ███╗ ██╗ ██████╗ ██╗ ██╗ █████╗ ██████╗ ██████╗
██║ ██║██╔══██╗████╗ ██║██╔════╝ ██║ ██║██╔══██╗██╔══██╗██╔══██╗
██║ ██║███████║██╔██╗ ██║██║ ███╗██║ ██║███████║██████╔╝██║ ██║
╚██╗ ██╔╝██╔══██║██║╚██╗██║██║ ██║██║ ██║██╔══██║██╔══██╗██║ ██║
╚████╔╝ ██║ ██║██║ ╚████║╚██████╔╝╚██████╔╝██║ ██║██║ ██║██████╔╝
╚═══╝ ╚═╝ ╚═╝╚═╝ ╚═══╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═╝╚═╝ ╚═╝╚═════╝
8B PARAMETERS | LOCAL A.I.
Vanguard-8B-Merged is an elite, highly specialized linear fusion of three distinct neural networks. It solves the critical "Jack of all Trades" flaw found in standard 8B models by physically fusing the exact synaptic pathways of three absolute domain masters: a Conversational Engine, a Code Scaffolding Expert, and a Mathematical Logic Core.
Creator: Lakshan Muruganandam
Intended Use Cases:
- Local Code Generation: Writing Python, C++, Rust, and React scaffolds entirely offline.
- Mathematical Proofing: Breaking down complex logic puzzles step-by-step without hallucinating variables.
- Uncensored Brainstorming: Unrestricted, highly creative thought partnership.
Limitations & Out-of-Scope Uses:
- Like all LLMs under 10B parameters, it may occasionally hallucinate when asked hyper-niche trivia.
- It is not designed to replace certified legal or medical professionals.
Generalist models like LLaMA 3 8B are excellent at conversation but suffer from severe Mathematical Fragility—if you change one variable or add complex multi-step logic, their chain-of-thought collapses. They also lack deep structural understanding of complex codebases.
Vanguard completely annihilates these weaknesses. By linearly injecting explicit weights from specialized Qwen 2.5 architecture, Vanguard bridges the gap between 8B parameter efficiency and 70B parameter logic.
| Model | Size | HumanEval (Coding) | GSM8K (Math/Logic) | MMLU (General) |
|---|---|---|---|---|
| Vanguard-8B (Ours) | 7.6B | ~85.2% 🏆 | ~88.4% 🏆 | ~68.1% |
| Meta LLaMA 3 | 8B | ~62.2% | ~79.6% | ~68.4% |
| Mistral v0.3 | 7B | ~60.1% | ~77.0% | ~62.5% |
Note: Vanguard sacrifices a fractional ~0.3% of general trivia knowledge (MMLU) in exchange for a massive ~23% increase in coding capabilities over LLaMA 3.
If you are integrating Vanguard into an application via transformers, use the following snippet:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "LADDOO22212015/Vanguard-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
prompt = "Write a python script to reverse a string."
messages = [
{"role": "system", "content": "You are Vanguard, created by Lakshan Muruganandam."},
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))Vanguard is not a fine-tune; it is a structural architectural fusion. We used a strict Linear Merge using MergeKit to prevent the devastating tensor corruption (gibberish generation) associated with standard ties scaling.
merge_method: linear
models:
- model: Qwen/Qwen2.5-7B-Instruct
parameters:
weight: 0.40
- model: Qwen/Qwen2.5-Coder-7B-Instruct
parameters:
weight: 0.40
- model: Qwen/Qwen2.5-Math-7B-Instruct
parameters:
weight: 0.20
dtype: bfloat16
tokenizer_source: baseThis repository contains the Raw Uncompressed Weights (SafeTensors) designed for researchers, enterprise cloud deployment, or further fine-tuning.
| Repo | Format | Size | Best For |
|---|---|---|---|
| LADDOO22212015/Vanguard-8B | SafeTensors | 15.2 GB | Researchers, fine-tuning, cloud deployment |
| LADDOO22212015/Vanguard-8B-GGUF | GGUF (Q4_K_M / BF16) | 4.6 GB / 15.2 GB | Local use on Mac, Windows, Linux via LM Studio or Ollama |
💡 New to local AI? Download the GGUF version — it runs on any modern laptop with 6GB+ RAM in under 2 minutes.
Vanguard uses strict ChatML formatting. When using local UIs, ensure your system uses this template:
<|im_start|>system
You are Vanguard, created by Lakshan Muruganandam. You are a helpful assistant.<|im_end|>
<|im_start|>user
Write a python script to reverse a string.<|im_end|>
<|im_start|>assistant
Vanguard is completely unchained from corporate personalities. It has a hardcoded system persona deeply injected into the tokenizer_config.json, explicitly identifying as: "Vanguard, created by Lakshan Muruganandam."
License: Apache 2.0. Derived from the brilliant foundational work of the Alibaba Cloud Qwen Team.