synapse-slm

Synapse SLM — Personalized Offline AI Assistant

Built by Abhinav Tyagi
📄 GitHub • 🔗 LinkedIn • 🧠 Live Demo

---

What is Synapse SLM?

Synapse SLM is a QLoRA fine-tuned Llama-3.2-3B model optimized for:

• Hinglish (Hindi-English code-switching) conversations
• Offline, CPU-only inference via 4-bit GGUF quantization
• Context-aware responses via an offline RAG pipeline
• Persona-consistent, instruction-tuned behavior

This is not a wrapper around an API — it runs fully locally, inside ~4GB RAM, at ~45 tokens/sec on CPU.

---

Model Details

---

Key Innovations

1. QLoRA Fine-Tuning for Behavioral Shaping

Fine-tuned with rank=16, alpha=32 on 3,500+ Hinglish instruction samples. The training objective wasn't just language — it was behavioral engineering: teaching the model when to explain, when to commit, and how to handle Hindi-English code-switching naturally.

Fine-tuning doesn't just improve answers — it rewires behavior. The training signal determines what feels "safe" to the model: explain vs. hedge, commit vs. qualify, answer vs. avoid.

2. 4-bit GGUF Quantization + CPU Inference

Converted to GGUF format using llama.cpp. Achieves ~45 tokens/sec on CPU-only hardware within a ~4GB RAM footprint — making it deployable on any laptop without GPU.

3. Offline RAG Pipeline

Implements embedding-based retrieval for local PDF/TXT ingestion. Supports context-aware responses without any cloud API dependency — fully air-gapped.

4. Hinglish Code-Switching

Trained specifically on Hindi-English mixed language patterns. Handles natural Hinglish input without requiring language detection or preprocessing.

---

Behavioral Study: How Fine-Tuning Changes Model Behavior

During development, Abhinav Tyagi trained two variants from the same base model to study behavioral drift:

• Model A (Synapse) — optimized for clarity, explanation, and usefulness
• Model B (Reflection-Heavy) — trained to emphasize uncertainty, limits, and caution

Key finding: Same architecture. Same tokenizer. Same base weights. Only the training signal differed — yet the behavioral output was completely different.

Model B wasn't hallucinating. It was over-aligned. And still useless.

Alignment without usability collapses into abstraction. Reasoning without explanation helps no one.

This study shaped Synapse's training philosophy: reasoning must serve explanation, not replace it.

---

Usage

With llama.cpp (Recommended for CPU)

# Install llama.cpp
pip install llama-cpp-python

# Run inference
from llama_cpp import Llama

llm = Llama(
    model_path="synapse-slm-q4.gguf",
    n_ctx=2048,
    n_threads=8
)

response = llm(
    "Bhai, explain karo gradient descent kya hota hai",
    max_tokens=512,
    temperature=0.7
)
print(response['choices'][0]['text'])

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("Abhinav-Tyagi/synapse-slm")
model = AutoModelForCausalLM.from_pretrained(
    "Abhinav-Tyagi/synapse-slm",
    torch_dtype=torch.float16,
    device_map="auto"
)

inputs = tokenizer("Explain neural networks simply", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Docker (Full Offline Setup)

git clone https://github.com/abhinavtyagi466/synapse-slm
cd synapse-slm
docker compose up
# Access at http://localhost:7860

---

Training Details

Base Model    : meta-llama/Llama-3.2-3B-Instruct
Method        : QLoRA
LoRA Rank     : 16
LoRA Alpha    : 32
Dataset Size  : 3,500+ instruction pairs
Languages     : English + Hindi + Hinglish
Quantization  : 4-bit GGUF (llama.cpp)
Inference     : ~45 tokens/sec on CPU
RAM           : ~4GB footprint
Deployment    : Docker containerized
RAG           : Offline PDF/TXT ingestion via dense embeddings

---

Offline RAG Pipeline

Synapse includes a fully offline RAG system:

1. Ingestion — Drop any PDF or TXT file into the /docs folder
2. Embedding — Documents are chunked and embedded locally (no API calls)
3. Retrieval — At query time, top-k relevant chunks are retrieved via FAISS
4. Generation — Retrieved context is injected into the prompt before generation

No internet required. No API keys. Fully private.

---

Performance

---

About the Author

Abhinav Tyagi is an LLM Engineer specializing in fine-tuning, quantization, and deployment of production-ready AI systems. He also built:

• Synapse-124M — A 124M parameter transformer built from scratch with GQA, MoE, Sliding Window Attention, NTK-RoPE, SwiGLU, and a custom BPE tokenizer
• Synapse Wingman — A full agentic AI desktop assistant controlled via Telegram with vision, WhatsApp automation, and multi-step task execution
• Smart Contextual RAG Chatbot — Hybrid RAG with CoVe (Chain of Verification), multi-query generation, FAISS, reducing cloud API costs by ~40%
• Psywarp — Published research on multimodal cognitive AI framework for emotion and behavior modeling (DOI: 10.5281/zenodo.18182199)

📧 [email protected]
🐙 GitHub
💼 LinkedIn

---

License

MIT License — free to use, modify, and distribute with attribution.

---

"Building AI that works offline, works fast, and actually works."
— Abhinav Tyagi