Spin-80k

Spin-80k is a lightweight, 80k-parameter decoder-only language model built from scratch by Quantech to demonstrate custom Transformer architecture

---

Model Specifications

• Organization: Quantech
• Architecture: Custom Decoder-only Transformer
• Total Parameters: ~80,112
• Layers: 2
• Hidden Dimension: 48
• Attention Heads: 4
• Feed-Forward Dimension: 128
• Positional Encoding: Rotary Position Embeddings (RoPE)
• Normalization: RMSNorm
• Activation: SwiGLU
• Vocabulary: 512 Byte-Pair Encoding (BPE) tokens
• Context Length: 256 tokens

---

Quickstart

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Quantech/spin-80k"


tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True)
model.eval()

prompt = "<|im_start|>user\nWrite a short story about a dog.<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=50,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

print(tokenizer.decode(outputs[0]))