---
datasets:
- blesspearl/stackexchange-math-sample
language:
- en
library_name: transformers
license: mit
---

[Guide](https://medium.com/@rajatsharma_33357/fine-tuning-llama-using-lora-fb3f48a557d5)
# Fine-Tuned LLaMA 3.1 Model on Stack Exchange Math Dataset

This repository contains the fine-tuned LLaMA 3.1 model using LoRA on a dataset collected from Stack Exchange Math. The model is designed to answer mathematical questions in a manner similar to Stack Exchange responses.

## Model Details

- **Base Model:** [Meta-Llama-3.1-8B](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B)
- **Fine-Tuned Model:** [math-stackexchange](https://huggingface.co/blesspearl/math-stackexchange)
- **Dataset:** [stackexchange-math-sample](https://huggingface.co/datasets/blesspearl/stackexchange-math-sample)
- **Training Environment:**
  - Framework: PyTorch with Transformers
  - Platform: Google Colab
  - Hardware: 1 x T4 GPU (15GB)

## Data Preparation

The dataset used for fine-tuning includes 1000 samples collected from Stack Exchange Math. Each sample consists of a question and its accepted answer.

### Preprocessing

The data was preprocessed using the following steps:
1. Loading the dataset from Hugging Face.
2. Shuffling the dataset and selecting 1000 samples.
3. Formatting the data into a chat template suitable for training.

## Training Details

### Libraries and Dependencies

```python
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments, pipeline, logging
from peft import LoraConfig, PeftModel, prepare_model_for_kbit_training, get_peft_model
from google.colab import drive, userdata
import os, torch, wandb
from trl import SFTTrainer, setup_chat_format
from huggingface_hub import login
```

### Loading Data and Model

```python
model_name = "meta-llama/Meta-Llama-3.1-8B"
dataset_name = "blesspearl/stackexchange-math-sample"

torch_dtype = torch.float16
attn_implementation = "eager"
wandb.login(key=userdata.get("WANDB_API_KEY"))
run = wandb.init(
    project='Fine tunning LLama-3.1-8b on math-stack-exchange',
    job_type="training",
    anonymous="allow"
)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch_dtype,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    attn_implementation=attn_implementation
)

tokenizer = AutoTokenizer.from_pretrained(model_name)
model, tokenizer = setup_chat_format(model, tokenizer)
```

### LoRA Configuration

```python
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=['up_proj', 'down_proj', 'gate_proj', 'k_proj', 'q_proj', 'v_proj', 'o_proj']
)
model = get_peft_model(model, peft_config)
```

### Data Preparation

```python
dataset = load_dataset(dataset_name, split="all")
dataset = dataset.shuffle(seed=65).select(range(1000))

def format_chat_template(row):
    row_json = [{"role": "user", "content": row["question_body"]},
                {"role": "assistant", "content": row["accepted_answer"]}]
    row["text"] = tokenizer.apply_chat_template(row_json, tokenize=False)
    return row

dataset = dataset.map(format_chat_template, num_proc=4)
dataset = dataset.train_test_split(test_size=0.2)
dataset = dataset.remove_columns(["question_body", "accepted_answer"])
```

### Training Configuration

```python
training_arguments = TrainingArguments(
    output_dir="math-stackexchange",
    per_device_train_batch_size=1,
    per_device_eval_batch_size=1,
    gradient_accumulation_steps=2,
    optim="paged_adamw_32bit",
    num_train_epochs=1,
    evaluation_strategy="steps",
    eval_steps=0.2,
    logging_steps=1,
    warmup_steps=10,
    logging_strategy="steps",
    learning_rate=2e-4,
    fp16=False,
    bf16=False,
    group_by_length=True,
    report_to="wandb"
)

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    peft_config=peft_config,
    max_seq_length=512,
    dataset_text_field="text",
    tokenizer=tokenizer,
    args=training_arguments,
    packing=False,
)
trainer.train()
wandb.finish()
model.config.use_cache = True
```

## Model and Dataset

- **Model:** [math-stackexchange](https://huggingface.co/blesspearl/math-stackexchange)
- **Dataset:** [stackexchange-math-sample](https://huggingface.co/datasets/blesspearl/stackexchange-math-sample)

## Usage

To use the fine-tuned model for inference, you can load it using the Hugging Face Transformers library and pass in your data for querying.

### Example Code

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "blesspearl/math-stackexchange"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

def answer_question(question):
    inputs = tokenizer(question, return_tensors="pt")
    outputs = model.generate(**inputs)
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return answer

question = "What is the derivative of sin(x)?"
answer = answer_question(question)
print(answer)
```

## Conclusion

This documentation provides an overview of the fine-tuning process of the LLaMA 3.1 model using LoRA on the Stack Exchange Math dataset. The model and dataset are available on Hugging Face for further use and exploration.

For any questions or issues, feel free to open an issue on the [model repository](https://huggingface.co/blesspearl/math-stackexchange).