Model Card for dzhuj/llm-course-hw3-lora

Model Details

Model Description

This is the fintuned version of OuteAI/Lite-Oute-1-300M-Instruct on text sentiment classification task into 3 categoriest: positive, negative or neutral. Finetuneing was performed using LoRA.

How to Get Started with the Model

from safetensors.torch import load_file
from huggingface_hub import hf_hub_download

REPO_NAME = dzhuj/llm-course-hw3
model = AutoModelForCausalLM.from_pretrained(f"{REPO_NAME}-lora", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(f"{REPO_NAME}-lora")
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"

# Примените peft к модели
apply_peft_to_module(model, LinearWithLoRA, r=8, alpha=16, target_submodules=["k_proj", "v_proj"])
model.to(DEVICE)

path = hf_hub_download(f"{REPO_NAME}-lora", "model.safetensors")
state_dict = load_file(path)

model.load_state_dict(state_dict, strict=False)

LoRA_saved_model_accuracy = eval(model, dataset["test"], tokenizer)
print(f"Accuracy after LoRA training: {LoRA_saved_model_accuracy:.2f}")

Training Details

Training Data

The following dataset was used: cardiffnlp/tweet_eval Example:

text: @user Alciato: Bee will invest 150 million in January, another 200 in the Summer and plans to bring Messi by 2017"
label: 2
str_label: positive

Training Procedure

Training Hyperparameters

• lora_r = 8;
• lora_alpha = 16;
• target_submodules = ["k_proj", "v_proj"];
• batch_size = 16;
• lr = 5e-4;
• num_epochs = 1

Evaluation

Testing Data, Factors & Metrics

The following testing protocol:

@torch.inference_mode
def eval(model, dataset, tokenizer, show_conf_m=True, batch_size=100):
    """Evaluates the given model on the provided dataset.

    Parameters:
        model: The language model used for generating sentiment predictions.
        dataset: An iterable collection of examples, where each example is a dict with keys:
            - "input_ids": The input text message.
            - "str_label": The ground truth sentiment label (e.g., "positive" or "negative").

    Returns:
        float: The macro f1 score
    """
    name2idx = {v: k for k, v in IDX2NAME.items()}
    name2idx[""] = len(name2idx)
    ground_truth = []
    predicted = []

    for examples in tqdm(dataset.batch(batch_size)):
        input_ids = pad(list(map(torch.tensor, examples["input_ids"])), padding_value=tokenizer.pad_token_id).to(DEVICE)
        attention_mask = pad(list(map(lambda it: torch.ones(len(it)), examples["input_ids"])), padding_value=0).to(
            DEVICE
        )
        output_ids = model.generate(input_ids, attention_mask=attention_mask, max_new_tokens=16)
        shrinked_ids = output_ids[:, input_ids.shape[1] :]
        texts = tokenizer.batch_decode(shrinked_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)

        for i in range(len(examples["str_label"])):
            predicted_sentiment = postprocess_sentiment(texts[i])
            ground_truth.append(name2idx[examples["str_label"][i]])
            predicted.append(name2idx[predicted_sentiment])

    if show_conf_m:
        conf_m = confusion_matrix(ground_truth, predicted, labels=list(name2idx.values()))
        disp = ConfusionMatrixDisplay(conf_m, display_labels=list(name2idx.keys()))
        disp.plot()

    f1 = f1_score(ground_truth, predicted, labels=list(name2idx.values()), average="macro", zero_division=0.0)
    return f1

Results

• F1 before fintuning: 0.16;
• F1 after LoRA fintuning: 0.52;