---
# For reference on model card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
# Doc / guide: https://huggingface.co/docs/hub/model-cards
{
  "library_name": "peft",
  "pipeline_tag": "text-generation",
  "license": "llama3.2",
  "tags": [
    "LoRA",
    "QLoRA",
    "instruction-tuning",
    "text-classification",
    "peft",
    "transformers",
    "trl",
    "bitsandbytes",
    "base_model:adapter:meta-llama/Llama-3.2-1B"
  ],
  "base_model": "meta-llama/Llama-3.2-1B",
  "datasets": ["real-jiakai/arxiver-with-category"],
  "language": ["en"],
  "widget": [
    {
      "text": "Classify the text into ['cs.CL','cs.CV','cs.LG','hep-ph','quant-ph'] and return the answer as the exact text label.\ntext: Quantum entanglement in photonics\nlabel:"
    },
    {
      "text": "Classify the text into ['cs.CL','cs.CV','cs.LG','hep-ph','quant-ph'] and return the answer as the exact text label.\ntext: Vision transformer achieves state-of-the-art on ImageNet\nlabel:"
    }
  ]
}

---

# Model Card for LLM Instruction‑Tuning for Text Classification (LoRA + QLoRA)

<!-- Provide a quick summary of what the model is/does. -->

This repository provides code and configuration to fine‑tune a decoder‑only LLM (default: `meta-llama/Llama-3.2-1B`) for **instruction‑style text classification** using **LoRA/QLoRA**. Rather than training a task‑specific classifier head, the project formulates classification as a short instruction → answer generation task and evaluates by **exact string match** against the label. It includes simple training/inference scripts, a 5‑label arXiv‑style demo, and optional Amazon SageMaker entrypoints.

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->

This project instruction‑tunes a base, decoder‑only LLM with **LoRA adapters** loaded in **4‑bit NF4** precision for memory‑efficient training and inference. Supervised fine‑tuning is performed with TRL’s `SFTTrainer`. Prompts ask the model to “return the answer as the exact text label,” so predictions are decoded as plain text and compared by string match.

- **Developed by:** Amirhossein Yousefi (GitHub: `amirhossein-yousefi`)
- **Model type:** Decoder‑only LLM fine‑tuned with LoRA for single‑label text classification via instruction‑following
- **Language(s) (NLP):** English by default (demo dataset uses arXiv titles/abstracts); broader multilingual coverage depends on the chosen base model
- **License:** The repository itself does not include an explicit OSS license; the **base model** `meta-llama/Llama-3.2-1B` is governed by the **Llama 3.2 Community License**. You must accept and comply with Meta’s license to access and use the weights. 
- **Finetuned from model :** `meta-llama/Llama-3.2-1B` (configurable)

### Model Sources 

<!-- Provide the basic links for the model. -->

- **Repository:** https://github.com/amirhossein-yousefi/LLM-Instruction-Tuning-Text-Classification
- **Demo :** The repo includes an arXiv‑style 5‑label demo and example results; no hosted demo is provided.

## Uses

### Direct Use

- Fine‑tune LoRA adapters on your own CSV dataset for **single‑label text classification** (e.g., topic/category detection) using the provided `scripts/train.py`.
- Run inference/evaluation with `scripts/predict.py` to generate deterministic label strings and compute **accuracy**, **micro/macro F1**, a **classification report**, and a **confusion matrix**.
- Optional **Amazon SageMaker** utilities let you run managed training and deploy a real‑time endpoint with the LoRA adapters attached at load time.

### Downstream Use 

- Integrate the trained LoRA adapters into applications where explainable, instruction‑driven classification is helpful (e.g., routing, tagging, moderation). 
- Swap the base model (any compatible decoder‑only LLM on the Hugging Face Hub) and re‑train with the same prompt template.
- Extend label sets without architectural changes—only prompt/label lists need to be updated.

### Out-of-Scope Use

- **CPU‑only** training/inference with this repo as‑is (4‑bit `bitsandbytes` path expects NVIDIA CUDA GPUs).
- **Multi‑label** classification (comma‑separated outputs) is not implemented out of the box (listed as a roadmap idea).
- **Open‑domain generation** or safety‑critical decision‑making; this project focuses on label selection with short inputs.

## Bias, Risks, and Limitations

- Outputs mirror biases in the **training corpus** you provide and in the **base model**. If your labels or examples are imbalanced or ambiguous, the model may propagate that bias.
- Exact‑match decoding can be brittle to **tokenization/typo** effects—ensure labels are short, canonical strings and restrict the decoding space.
- The base Llama 3.2 model has its own safety limitations and license‑based usage constraints (e.g., attribution and acceptable‑use provisions). 
- The demo dataset is limited to **5 arXiv‑style labels** and relatively short academic texts; generalizing beyond this domain requires additional data.

### Recommendations

- Curate balanced datasets; consider **stratified splits** and per‑class metrics.
- Keep **temperature = 0.0** for deterministic label decoding; constrain generation length (e.g., `max_new_tokens=8`).
- Validate robustness with **label synonyms/aliases** and adversarial cases; consider post‑processing that maps variants to canonical labels.
- Review and comply with the **Llama 3.2 Community License** (and any other upstream licenses) when distributing adapters/derivatives.

## How to Get Started with the Model

**Install & train**

```bash
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\Activate.ps1
pip install --upgrade pip
pip install -r requirements.txt

# If the base model is gated, export an HF token
export HF_TOKEN=YOUR_HF_ACCESS_TOKEN

# One‑command training on CSVs
python scripts/train.py   --base_path dataset   --train_file train.csv   --val_file validation.csv   --test_file test.csv   --label_column label_name   --text_fields title abstract   --base_model_name meta-llama/Llama-3.2-1B   --output_dir llama-3.2-1b-arxiver-lora
```

**Inference & evaluation**

```bash
python scripts/predict.py   --base_path dataset   --test_file test.csv   --base_model_name meta-llama/Llama-3.2-1B   --output_dir llama-3.2-1b-arxiver-lora   --save_csv predictions.csv
```

**SageMaker **

```bash
# Train a managed job
python sagemaker/train_sm.py   --source_dir .   --dataset_dir dataset   --train_file train.csv   --val_file validation.csv   --test_file test.csv   --label_column label_name   --text_fields title abstract   --base_model_id meta-llama/Llama-3.2-1B   --instance_type ml.g5.2xlarge   --instance_count 1

# Deploy a real‑time endpoint
python sagemaker/deploy_sm.py   --training_job_name <your-job>   --base_model_id meta-llama/Llama-3.2-1B   --instance_type ml.g5.2xlarge   --default_labels_json '["cs.CL","cs.CV","cs.LG","hep-ph","quant-ph"]'
```

## Training Details

### Training Data

- Expected input: three CSV files under a base path: `train.csv`, `validation.csv`, `test.csv`.
- Required columns: a **label** column (default `label_name`) and one or more text fields (defaults: `title`, `abstract`). Missing/blank text fields are skipped; text fields are concatenated with punctuation.
- The repository ships utilities to prepare a **5‑class arXiv‑style demo** (labels: `['cs.CL','cs.CV','cs.LG','hep-ph','quant-ph']`).

### Training Procedure

#### Preprocessing 

- Prompts are constructed as short instruction → answer pairs:
  - **Train:** includes the gold label after `label:`.
  - **Inference:** leaves `label:` empty and decodes the generated label.

#### Training Hyperparameters

- **Training regime:** mixed precision with `fp16=True`, `tf32=True`; 4‑bit NF4 quantization with bfloat16 compute (QLoRA‑style).
- **Selected defaults (single‑GPU):**
  - `num_train_epochs=1`
  - `per_device_train_batch_size=8`, `per_device_eval_batch_size=8`
  - `gradient_accumulation_steps=2` (effective 16 per step, per device)
  - `learning_rate=2e-4`, `weight_decay=1e-3`, `warmup_ratio=0.03`
  - `logging_steps=10`, `evaluation_strategy="epoch"`, `save_strategy="epoch"`, `save_total_limit=2`
  - LoRA: `r=2`, `alpha=2`, `dropout=0.0`
  - Quantization: `load_in_4bit=True`, `bnb_4bit_quant_type="nf4"`, `bnb_4bit_compute_dtype="bfloat16"`, `bnb_4bit_use_double_quant=True`
  - Generation (eval): `temperature=0.0`, `max_new_tokens=8`, `do_sample=False`

#### Speeds, Sizes, Times 

- Example environment: Laptop RTX 3080 Ti (16 GB VRAM), CUDA 12.9, PyTorch 2.8.0+cu129.
- Example run stats: ~6,314 seconds wall‑clock training, with TensorBoard logs under the run directory.
- Total training FLOPs (example): ~3.69e16 (as reported by the training logs).

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data

- The example evaluation uses the provided arXiv‑style 5‑label test split.

#### Factors

- Per‑class metrics are reported for `cs.CL`, `cs.CV`, `cs.LG`, `hep-ph`, `quant-ph`.

#### Metrics

- Accuracy, micro F1, macro F1, per‑class precision/recall/F1, and a confusion matrix.

### Results

- **Overall:** Accuracy 93.8%, Micro‑F1 0.938, Macro‑F1 0.950.
- **Per‑class (Precision / Recall / F1 / Support):**
  - `cs.CL`: 0.914 / 0.963 / 0.938 / 432
  - `cs.CV`: 0.935 / 0.923 / 0.929 / 545
  - `cs.LG`: 0.917 / 0.890 / 0.903 / 536
  - `hep-ph`: 0.994 / 0.988 / 0.991 / 164
  - `quant-ph`: 0.986 / 0.990 / 0.988 / 293

#### Summary

The LoRA‑tuned 1B parameter Llama 3.2 model achieves strong performance on short academic texts while keeping training/inference affordable due to 4‑bit quantization. Performance is consistent across most classes, with particularly high scores for physics categories.

## Model Examination 

- The repo includes utilities for a **classification report** and **confusion matrix**. Inspect misclassifications to refine label definitions or add examples. Consider probing sensitivity to prompt wording.

## Environmental Impact

*(Approximate; depends on your hardware and run length.)*  
Use the [MLCO2 Impact calculator](https://mlco2.github.io/impact#compute) with your GPU model, power draw, and wall‑clock runtime.

- **Hardware Type:** Single NVIDIA GPU (example: RTX 3080 Ti Laptop 16 GB)
- **Hours used:** ~1.75 hours (example)

## Technical Specifications 

### Model Architecture and Objective

- **Architecture:** Decoder‑only Transformer (Llama 3.2 family when using the default base)
- **Objective:** Supervised instruction‑tuning for **single‑label classification** via generative decoding with exact‑match evaluation
- **Context length:** 512 tokens (config default; pass explicitly to trainer to ensure enforcement)

### Compute Infrastructure

#### Hardware

- NVIDIA CUDA GPU required for 4‑bit `bitsandbytes` training/inference  
  (CPU‑only runs are not supported by the included scripts).

#### Software

- Python ≥ 3.10, PyTorch, `transformers`, `trl`, `peft`, `bitsandbytes`, `accelerate`, and standard scientific Python packages.
- Optional: Astral’s `uv` for faster, reproducible dependency management (the repo also ships `requirements.txt`).

## Citation 

If you use this repository, please cite the GitHub project and the base model as appropriate.

**BibTeX (project):**
```bibtex
@software{yousefi_2025_llm_instruction_tuning_text_classification,
  author    = {Yousefi, Amirhossein},
  title     = {LLM Instruction-Tuning for Text Classification (LoRA + QLoRA)},
  year      = {2025},
  publisher = {GitHub},
  url       = {https://github.com/amirhossein-yousefi/LLM-Instruction-Tuning-Text-Classification}
}
```

**APA (project):**  
Yousefi, A. (2025). *LLM Instruction‑Tuning for Text Classification (LoRA + QLoRA)*. GitHub. https://github.com/amirhossein-yousefi/LLM-Instruction-Tuning-Text-Classification

**Base model:** Meta AI. (2024). *Llama 3.2‑1B* [Computer software]. Meta. https://huggingface.co/meta-llama/Llama-3.2-1B

## Glossary 

- **LoRA:** Low‑Rank Adapters for parameter‑efficient fine‑tuning.
- **QLoRA:** LoRA training with quantized base weights (typically 4‑bit NF4) and higher‑precision compute.
- **SFT:** Supervised Fine‑Tuning.
- **Exact‑match decoding:** Evaluates whether the generated label text exactly matches the gold label string.

## More Information 

- Amazon SageMaker scripts are included for managed training and deployment.
- Roadmap ideas include multi‑label support and few‑shot exemplars in prompts.

## Model Card Authors 

- Drafted by: ChatGPT (based on the repository’s README and code structure)
- Repository author: Amirhossein Yousefi

## Model Card Contact

- Open an issue on the GitHub repository for questions or contributions.