This is a LoRA adapter for Qwen/Qwen2.5-0.5B-Instruct, fine-tuned with QLoRA for binary sentiment classification on the Stanford IMDB movie review dataset.

This adapter is part of an end-to-end MLOps portfolio project demonstrating the full lifecycle of LLM fine-tuning: data preparation, QLoRA training, experiment tracking with MLflow, evaluation, and deployment.

Model Details

Base model: Qwen2.5-0.5B-Instruct
Fine-tuning method: LoRA (rank=8, alpha=16, dropout=0.05) on top of a 4-bit quantized base model (QLoRA)
Target modules: q_proj, k_proj, v_proj, o_proj
Trainable parameters: ~1.08M (0.22% of total model parameters)
Task: Binary sentiment classification (positive/negative), framed as an instruction-following text generation task
Training hardware: Single NVIDIA T4 GPU (Kaggle Notebooks, free tier)

Training Data

Dataset: Stanford IMDB (stanfordnlp/imdb), 25,000 labeled movie reviews in the original train split
Split: 90% train / 10% validation, stratified by label, seed=42
Test set: The original 25,000-example IMDB test split was held out entirely and never used during training or hyperparameter selection

Training Procedure

Prompt format: Chat-style conversation (system / user / assistant) using Qwen2.5's native chat template
Loss masking: Loss computed only on the assistant response tokens (assistant_only_loss=True); system and user tokens are masked
Sequence length: Reviews truncated to 150 words before tokenization to guarantee the response is never cut off by the 256-token sequence limit
Epochs: 1
Effective batch size: 16
Learning rate: 2e-4, cosine schedule with warmup
Precision: FP16 throughout (base model, LoRA weights, and training loop) — chosen for compatibility with the T4 GPU's Turing architecture, which lacks native BF16 tensor core support

Evaluation Results

Evaluated on a stratified sample of 1,000 examples from the held-out IMDB test set (seed=42), using true autoregressive generation (no teacher forcing) — the model receives only the system and user turns and must generate its own response.

Metric	Value
Accuracy	0.9176
Precision	0.9314
Recall	0.9014
F1	0.9162

Confusion matrix (995 valid responses; 5 out of 1,000 generations were not recognized as either label):

	Predicted: negative	Predicted: positive
True: negative	465	33
True: positive	49	448

How to Use

python

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

BASE_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
ADAPTER_REPO = "Carlo88/qwen05b-sentiment-lora-imdb"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    dtype=torch.float16,
)

model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)
model.eval()

review = "This movie was a complete waste of time, the plot made no sense."
messages = [
    {"role": "system", "content": "Sei un classificatore di sentiment. Rispondi solo con 'positive' o 'negative'."},
    {"role": "user", "content": review},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=5, do_sample=False)
response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)  # -> "negative"

Limitations

The system prompt used during training is in Italian; the adapter has only been validated with that exact system prompt. Using a different system prompt (or language) may degrade performance.
Reviews are truncated to the first 150 words before classification; very long reviews may lose relevant context from later paragraphs.
The raw IMDB dataset contains unescaped HTML artifacts (e.g. <br /><br />) which were not cleaned during preprocessing. This did not appear to harm performance but is worth noting for anyone extending this work.
Evaluated on a stratified sample of the test set (1,000 of 25,000 examples), not the full test set.

Training Framework & Tools

Hugging Face transformers, peft, trl, bitsandbytes
Experiment tracking: MLflow
Training environment: Kaggle Notebooks (free T4 GPU)

Citation

If you use this adapter, please also cite the base model and dataset:

Qwen2.5: Qwen Team, Alibaba Group
Stanford IMDB: Maas et al., 2011, Learning Word Vectors for Sentiment Analysis