app.py

from unsloth import FastLanguageModel
import torch
import pandas as pd
from datasets import Dataset
import numpy as np
from sklearn.model_selection import train_test_split
from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

max_seq_length = 4096 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.


model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/tinyllama-bnb-4bit", # "unsloth/tinyllama" for 16bit loading
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 32, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 32,
    lora_dropout = 0, # Currently only supports dropout = 0
    bias = "none",    # Currently only supports bias = "none"
    use_gradient_checkpointing = False, # @@@ IF YOU GET OUT OF MEMORY - set to True @@@
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

alpaca_prompt = """Below is an instruction that describes a task, paired with an output that provides correct output for that task. Write a response that produces correct solution to the problem

### Instruction:
{}

### Input:
{}

### Response:
{}"""

EOS_TOKEN = tokenizer.eos_token
def formatting_prompts_func(examples):
    instructions = "The problem has the following answer. Understand step-by-step how it is solved to produce the correct solution and then produce the correct solution"
    inputs       = examples["Riddle"]
    outputs      = examples["Answer"]
    texts = []
    for instruction, input, output in zip(instructions, inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }



df = pd.read_csv('math_riddles.csv')
train, test = train_test_split(df, test_size=0.2, random_state=42)
train_ds = Dataset.from_pandas(train)
test_ds = Dataset.from_pandas(test)

tokenized_train = train_ds.map(formatting_prompts_func, batched=True,
                               remove_columns=['Riddle', 'Answer', '__index_level_0__']) # Removing features

tokenized_test = test_ds.map(formatting_prompts_func, batched=True,
                               remove_columns=['Riddle', 'Answer']) # Removing features



trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = tokenized_train,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 24,
    packing = True, # Packs short sequences together to save time!
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 1,
        warmup_ratio = 0.1,
        num_train_epochs = 3,
        learning_rate = 2e-5,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.1,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "none", # Use this for WandB etc
    ),
)

trainer_stats = trainer.train()

# Define inference function
def inference(instruction, user_input):
    prompt = alpaca_prompt.format(
        instruction,
        user_input,
        ""  # Leave output blank for generation
    )

    inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

    outputs = model.generate(
        **inputs,
        max_new_tokens=64,
        use_cache=True
    )

    # Fix: Define result before printing it
    result = tokenizer.batch_decode(outputs)[0]
    print(result)  # Now you can print it

    # Extract just the generated response (after the prompt)
    response_prefix = "### Response:"
    if response_prefix in result:
        result = result.split(response_prefix)[1].strip()

    return result

# Create Gradio interface
import gradio as gr
demo = gr.Interface(
    fn=inference,
    inputs=[
        gr.Textbox(label="Instruction", value="Solve the problem"),
        gr.Textbox(label="Input", value="There is a three digit number.The second digit is four times as big as the third digit, while the first digit is three less than the second digit.What is the number?")
    ],
    outputs="text",
    title="Language Model Interface",
    description="Enter an instruction and input to generate a response from the model."
)


demo.launch(share=True)