SAWiT AI Hackathon: Colloquial Dataset Creation and ModelTraining stringlengths 1 28.5k | Unnamed: 1 stringclasses 2
values | Unnamed: 2 stringclasses 2
values |
|---|---|---|
Welcome to the SAWiT AI Hackathon! In thishackathon, you'll be creating a colloquial language dataset for one ofthe following languages: | null | null |
Marathi | null | null |
Tamil | null | null |
Hindi | null | null |
Telugu | null | null |
Malayalam | null | null |
Bengali | null | null |
Task Overview | null | null |
1. Pick a Language: Choose one language from the listabove. | null | null |
2. Create a Colloquial Dataset: Collect and curate adataset that represents the colloquial language of your selectedlanguage. This could include informal conversations, slang, and regionalvariations. | null | null |
3. Train a Model: Use your dataset to train a machinelearning model that understands or processes the colloquial language.You can use existing models and fine-tune them with your dataset. | null | null |
4. Push the Dataset & Model to Hugging Face: Oncethe model is trained, push both your dataset and model to Hugging Face. You will need to createa Hugging Face account if you don't have one already. | null | null |
5. Share the Final Links for Evaluation: After pushingthe dataset and model to Hugging Face, share the final links forevaluation in HackathonPlatform. | null | null |
Helpful Resources | null | null |
Unsloth: Anoptimization framework for fine-tuning Large Language Models (LLMs) thatmakes training 2-4x faster. It provides optimized implementations ofcommon operations like LoRA (Low-Rank Adaptation) training, specializedkernels for faster computation, and memory-efficient training methods.The framework integrates wit... | null | null |
Hugging Face:A platform for hosting and sharing your datasets and models. You canpush your trained models and datasets to Hugging Face and share the linkfor evaluation. | null | null |
Steps Overview: | null | null |
1. Dataset Creation: Use Unsloth or similar approachfor dataset creation (link above). | null | null |
2. Model Training: Fine-tune or train a model usingyour dataset. | null | null |
3. Push to Hugging Face: Upload your model and datasetto Hugging Face. | null | null |
4. Evaluation: Share your Hugging Face model URL forevaluation. | null | null |
Example Output | null | null |
`` | null | null |
Input to Translate: "What is data structure?" | null | null |
Output Expected (Tamil): "Data structure nayenna?" | null | null |
`` | null | null |
Good Luck! | null | null |
We are excited to see your contributions! Happy coding and training!🚀 | null | null |
Below is an example that demonstrates the process of sample dataset creation, model training using Unsloth, and uploading it to Hugging Face | null | null |
# Install required packages!pip install torch==2.5.1!pip install transformers datasets accelerate bitsandbytes!pip install unsloth!pip install peftimport torchfrom datasets import Datasetfrom unsloth import FastLanguageModelimport pandas as pdfrom datetime import datetimefrom transformers import TrainingArguments, Trai... | null | null |
Requirement already satisfied: torch==2.5.1 in /usr/local/lib/python3.11/dist-packages (2.5.1+cu124)Requirement already satisfied: filelock in /usr/local/lib/python3.11/dist-packages (from torch==2.5.1) (3.17.0)Requirement already satisfied: typing-extensions>=4.8.0 in /usr/local/lib/python3.11/dist-packages (from torc... | null | null |
{"id":"616b368a18db4e1a84dfd5d3a0e2cf77","pip_warning":{"packages":["google"]}} | null | null |
Requirement already satisfied: peft in /usr/local/lib/python3.11/dist-packages (0.14.0)Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.11/dist-packages (from peft) (1.26.4)Requirement already satisfied: packaging>=20.0 in /usr/local/lib/python3.11/dist-packages (from peft) (24.2)Requirement already... | null | null |
# Initialize model with unsloth and PEFTfrom peft import LoraConfig, get_peft_model# First initialize the modelMODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"MAX_LENGTH = 128# Initialize base modelmodel, tokenizer = FastLanguageModel.from_pretrained( model_name=MODEL_NAME, max_seq_length=MAX_LENGTH, dtype=to... | null | null |
Unsloth: WARNING `trust_remote_code` is True.Are you certain you want to do remote code execution?==((====))== Unsloth 2025.2.12: Fast Llama patching. Transformers: 4.48.3. \\ /| GPU: Tesla T4. Max memory: 14.741 GB. Platform: Linux.O^O/ \_/ \ Torch: 2.6.0+cu124. CUDA: 7.5. CUDA Toolkit: 12.4. Triton: 3.2.0\ ... | null | null |
Device does not support bfloat16. Will change to float16. | null | null |
trainable params: 4,505,600 || all params: 1,104,553,984 || trainable%: 0.4079 | null | null |
# Create training dataset, this is an example, You can load from hugging facetraining_data = { 'text': [ "### Human: Translate to Tamil colloquial: What is data structure?\n### Assistant: Data structure na yenna?", "### Human: Translate to Tamil colloquial: How are you doing?\n### Assistant: Eppadi iru... | null | null |
{"model_id":"dd733ba56e1848d9963f358ad39dc599","version_major":2,"version_minor":0} | null | null |
{"model_id":"776f495a5a7e46ebb4dfe4f86ce9a5a5","version_major":2,"version_minor":0} | null | null |
Training examples: 8Validation examples: 2 | null | null |
# Hugging Face User Namehugging_face_user_name="Samhithaeleti"# Training configurationtraining_args = TrainingArguments( output_dir=f"./english-tamil-colloquial-{datetime.now().strftime('%Y%m%d-%H%M%S')}", per_device_train_batch_size=4, per_device_eval_batch_size=4, num_train_epochs=10, learning_rate=3e-... | null | null |
/usr/local/lib/python3.11/dist-packages/transformers/training_args.py:1575: FutureWarning: `evaluation_strategy` is deprecated and will be removed in version 4.46 of 🤗 Transformers. Use `eval_strategy` instead warnings.warn( | null | null |
# Login to Hugging Facefrom huggingface_hub import notebook_loginnotebook_login() | null | null |
{"model_id":"1faea425448c49a3906b9dc8d37f7b9e","version_major":2,"version_minor":0} | null | null |
# Initialize trainer and start training (replace the existing training code)trainer = Trainer( model=model, args=training_args, train_dataset=tokenized_train, eval_dataset=tokenized_val)# Modified training with progress barprint("Starting training...")trainer.train()# Save and push to hubtrainer.save_model(... | null | null |
Starting training... | null | null |
==((====))== Unsloth - 2x faster free finetuning | Num GPUs = 1 \\ /| Num examples = 8 | Num Epochs = 10O^O/ \_/ \ Batch size per device = 4 | Gradient Accumulation steps = 2\ / Total batch size = 8 | Total steps = 10 "-____-" Number of trainable parameters = 4,505,600 | null | null |
[10/10 00:08, Epoch 10/10] | null | null |
Step | Training Loss | Validation Loss |
2 | 14.906500 | 9.325418 |
4 | 14.906500 | 9.325418 |
6 | 14.906500 | 9.325418 |
8 | 14.906500 | 9.325418 |
10 | 14.906500 | 9.325418 |
Unsloth: Not an error, but LlamaForCausalLM does not accept `num_items_in_batch`.Using gradient accumulation will be very slightly less accurate.Read more on gradient accumulation issues here: https://unsloth.ai/blog/gradientNo files have been modified since last commit. Skipping to prevent empty commit.WARNING:hugging... | null | null |
Preparing model for inference... | null | null |
def translate_to_tamil_colloquial(text): prompt = f"""### Human: You are a Tamil colloquial language translator. Translate the following English text to Tamil colloquial language (spoken Tamil).Here are some examples:"What is this?" -> "Idhu enna?""How are you?" -> "Eppadi irukka?""Where are you going?" -> "Enga por... | null | null |
Testing the model:
English: explain deep learning?Tamil Colloquial: Deep Learning refers to an approach that learns complex patterns from data, which allows computers to understand and learn information beyond its--------------------------------------------------English: What is data structure?Tamil Colloquial: 1) A st... | null | null |
Interactive Demo | null | null |
Try your own sentences below: | null | null |
from IPython.display import HTML, displaydef interactive_translation(text_input): if text_input: translation = translate_to_tamil_colloquial(text_input) display(HTML(f""" <div style='padding: 10px; border-radius: 5px;'> <b>English:</b> {text_input}<br> <b>Tamil Colloquial:<... | null | null |
English: how are you | null | null |
Tamil Colloquial: I am doing fine, thank you! I am happy to hear that you have also been enjoying your trip. How about | null | null |
README.md exists but content is empty.
- Downloads last month
- 3