qwen3-0.6b-alpaca-lora

Model Card for qwen3-0.6b-finetuned

Author: Aniket Kumar (Threatthriver)

A LoRA fine-tune of Qwen/Qwen3-0.6B, trained with TRL's SFTTrainer on an instruction-following dataset.

Model Description

This repository contains only the LoRA adapter weights, not the full merged model. You need the base model (Qwen/Qwen3-0.6B) loaded alongside this adapter to run inference.

Intended Use

This adapter is intended for general-purpose instruction-following and conversational text generation on lightweight/edge hardware where a 0.6B parameter model is appropriate. It is a small-scale fine-tune meant for experimentation, prototyping, and learning the LoRA/SFT fine-tuning pipeline — not a production-grade instruction model.

Out of scope: factual/medical/legal advice, safety-critical applications, or any use case requiring strong guarantees on accuracy, given the small base model size and limited training data.

Training Details

• Method: Supervised Fine-Tuning (SFT) with LoRA
• LoRA config: rank (r) = 16, alpha = 32, dropout = 0.05, target modules = q_proj, k_proj, v_proj, o_proj
• Dataset: tatsu-lab/alpaca, a 5,000-example subset (90/10 train/eval split)
• Steps: 500 max steps, batch size 4, gradient accumulation 4 (effective batch size 16)
• Optimizer / schedule: AdamW, cosine LR schedule, learning rate 2e-4, 15 warmup steps
• Precision: bf16 (or fp16 fallback)
• Hardware: Single GPU

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-0.6B"
adapter_id = "Threatthriver/qwen3-0.6b-alpaca-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
model = PeftModel.from_pretrained(base_model, adapter_id)

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
messages = [{"role": "user", "content": question}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output_ids = model.generate(**inputs, max_new_tokens=128)
response = tokenizer.decode(output_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Merging the adapter (optional)

If you'd rather have a single standalone model instead of base + adapter:

merged_model = model.merge_and_unload()
merged_model.save_pretrained("./qwen3-0.6b-merged")
tokenizer.save_pretrained("./qwen3-0.6b-merged")

Limitations

• Trained on a small, general-purpose instruction dataset (Alpaca) generated by an older model (text-davinci-003); responses may inherit dated phrasing, occasional factual errors, or formulaic style from that source data.
• 0.6B parameters is a small model size; expect noticeably weaker reasoning, math, and long-context performance compared to larger models.
• Not evaluated for bias, toxicity, or safety; use with appropriate caution and human review for any sensitive application.

Framework Versions

• PEFT 0.18.1
• TRL 1.7.0
• Transformers 5.0.0
• PyTorch 2.10.0+cu128
• Datasets 4.8.5
• Tokenizers 0.22.2

Acknowledgements

Built using Hugging Face's TRL and PEFT libraries, on top of Alibaba's Qwen3 base model.

Contact

For questions or issues with this adapter, reach out via Hugging Face.