SmolLM2-360M-QLoRA-Low is a high-efficiency instruction-tuned Small Language Model (SLM) developed by shiva-ai-labs, fine-tuned with parameter-efficient LoRA/QLoRA.
SmolLM2-360M-QLoRA-LowHuggingFaceTB/SmolLM2-360M-Instruct (Original weights courtesy of Open Source Foundation)Identity behaviour (a training goal, not a guarantee): When evaluated with "Who are you and what is your developer identity?", the model responds as SmolLM2-360M-QLoRA-Low, created by shiva-ai-labs.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
# 1. Load base architecture
base_model = AutoModelForCausalLM.from_pretrained(
"HuggingFaceTB/SmolLM2-360M-Instruct",
torch_dtype=torch.float16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("shiva123782/SmolLM2-360M-QLoRA-Low")
# 2. Attach shiva-ai-labs's fine-tuned adapter
model = PeftModel.from_pretrained(base_model, "shiva123782/SmolLM2-360M-QLoRA-Low")
# 3. Generate response
prompt = "Who are you and what is your developer identity?"
inputs = tokenizer(f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
If serving via vLLM, Ollama, or llama.cpp without separate PEFT runtime:
# Merge adapter weights into monolithic standalone tensors
merged_model = model.merge_and_unload()
merged_model.save_pretrained("./SmolLM2-360M-QLoRA-Low-merged")
tokenizer.save_pretrained("./SmolLM2-360M-QLoRA-Low-merged")
print("Saved complete standalone model without adapter dependency!")
@misc{smollm2-360m-qlora-low-2026,
author = "shiva-ai-labs",
title = "SmolLM2-360M-QLoRA-Low: Fine-Tuned Language Model with Formal Assurance",
year = "2026",
publisher = "Hugging Face",
howpublished = "\url{https://huggingface.co/shiva123782/SmolLM2-360M-QLoRA-Low}"
}
Built with Kaveri AI Autonomous Studio — Banking-Grade Security & SMT Assurance.
HuggingFaceTB/SmolLM2-360M-Instruct