Full-Fine-Tuning-Qwen-2.5-0.5B-instruct-sft-prefix-tuned

Full-Fine-Tuning-Qwen-2.5-0.5B-instruct-sft-prefix-tuned

Overview

This repository contains a Prefix Tuning (PEFT) adapter trained on top of my Supervised Fine-Tuned (SFT) personalized Qwen model.

Instead of updating all model weights, Prefix Tuning learns a small set of trainable prefix key/value vectors while keeping the underlying language model completely frozen. This significantly reduces the number of trainable parameters while retaining the personalized knowledge learned during supervised fine-tuning.

Note: This repository contains only the Prefix Tuning adapter, not the complete language model. To use this adapter, you must first load the corresponding SFT model.

---

Base Model

This Prefix Tuning adapter was trained on the following model:

BASE_MODEL = "vishnuamarapu/Full-Fine-Tuning-Qwen-2.5-0.5B-instruct-sft"

Do not load the original Qwen/Qwen2.5-0.5B-Instruct model directly, as the adapter was trained specifically on the personalized SFT model.

---

Installation

pip install -U transformers peft accelerate torch

---

Loading the Prefix Tuning Adapter

import torch

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

BASE_MODEL = "vishnuamarapu/Full-Fine-Tuning-Qwen-2.5-0.5B-instruct-sft"

PREFIX_ADAPTER = (
    "vishnuamarapu/"
    "Full-Fine-Tuning-Qwen-2.5-0.5B-instruct-sft-prefix-tuned"
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype="auto"
)

model = PeftModel.from_pretrained(
    model,
    PREFIX_ADAPTER
)

---

Example Inference

messages = [
    {
        "role": "system",
        "content": (
            "You are Vishnu's personal AI assistant. "
            "Answer questions about Vishnu."
        )
    },
    {
        "role": "user",
        "content": "What is your name?"
    }
]

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=80,
    do_sample=False
)

response = tokenizer.decode(
    outputs[0],
    skip_special_tokens=True
)

print(response)

---

Training Method

• Base Architecture: Qwen2.5-0.5B-Instruct
• Base Personalized Model: Full Supervised Fine-Tuning (SFT)
• PEFT Method: Prefix Tuning
• Backbone Parameters: Frozen
• Trainable Parameters: Prefix Key/Value Vectors Only
• Task: Personalized Question Answering

---

Repository Contents

adapter_config.json
adapter_model.safetensors
chat_template.jinja
tokenizer.json
tokenizer_config.json
training_args.bin

The repository stores only the Prefix Tuning adapter, making it lightweight and efficient to distribute.

---

Intended Use

This adapter is intended for:

• Personalized AI assistants
• Parameter-Efficient Fine-Tuning (PEFT) research
• Prefix Tuning experiments
• Educational and learning purposes
• Benchmarking against other PEFT techniques such as Prompt Tuning, LoRA, and QLoRA

---

Notes

• This adapter must be used together with the SFT model listed above.
• It was trained specifically for answering questions about Vishnu Amarapu.
• The adapter is lightweight and contains only the learned Prefix Tuning parameters, making it much smaller than a fully fine-tuned model.

---

License

Please follow the license terms of the underlying Qwen model when using this adapter.