--- language: - en - te license: llama2 base_model: abhinand/dr-llama-te-instruct-v0 model-index: - name: telugu-llama-instruct-v0.1 results: [] --- # Telugu LLaMA 7B Instruct v0.1 Welcome to the inaugural release of the Telugu LLaMA 7B instruct model – an important step in advancing LLMs for the Telugu language. This model is ready for immediate inference and is also primed for further fine-tuning to cater to your specific NLP tasks. To dive deep into the development and capabilities of this model, please read the [research paper](https://arxiv.org/abs/2311.05845) and the [introductory blog post (WIP)]() that outlines our journey and the model's potential impact. > **Note:** This model is based on the Tamil LLaMA series of models. The GitHub repository remains the same - [https://github.com/abhinand5/tamil-llama](https://github.com/abhinand5/tamil-llama). The base models and the updated code for Tamil LLaMA v0.2 (which this work is based on) will be released soon. If you appreciate this work and would like to support its continued development, consider [buying me a coffee](https://www.buymeacoffee.com/abhinand.b). Your support is invaluable and greatly appreciated. [!["Buy Me A Coffee"](https://www.buymeacoffee.com/assets/img/custom_images/orange_img.png)](https://www.buymeacoffee.com/abhinand.b) ## Demo: To access an easy-to-use, no-code demo, please open the provided Google Colab notebook. Complete instructions for usage are included within the notebook itself. Demo In Colab ## Model description The Telugu LLaMA models have been enhanced and tailored specifically with an extensive Telugu vocabulary of ~16,000 tokens, building upon the foundation set by the original LLaMA-2. - **Model type:** A 7B parameter GPT-like model finetuned on ~500,000 samples consisting of an equal proportion of English and Telugu samples. (Dataset will be released soon) - **Language(s):** Bilingual. English and Telugu. - **License:** GNU General Public License v3.0 - **Finetuned from model:** [To be released soon]() - **Training Precision:** `bfloat16` - **Code:** [GitHub](https://github.com/abhinand5/tamil-llama) (To be updated soon) ## Prompt Template: ChatML ``` <|im_start|>system {system_message}<|im_end|> <|im_start|>user {prompt}<|im_end|> <|im_start|>assistant ``` ## Benchmark Results Benchmarking was done using [LLM-Autoeval](https://github.com/mlabonne/llm-autoeval) on an RTX 3090 on [runpod](https://www.runpod.io/). > **Note:** Please note that discrepancies have been observed between the Open LLM Leaderboard scores and those obtained from local runs using the LM Eval Harness with identical configurations. The results mentioned here are based on our own benchmarking. To replicate these findings, you can utilize the LLM-Autoeval or use [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) locally with the configurations described in Open LLM Leaderboard's About page. | Benchmark | Llama 2 Chat | Tamil Llama v0.2 Instruct | Telugu Llama Instruct | Malayalam Llama Instruct | |---------------|--------------|---------------------------|-----------------------|--------------------------| | ARC Challenge (25-shot) | 52.9 | **53.75** | 52.47 | 52.82 | | TruthfulQA (0-shot) | 45.57 | 47.23 | **48.47** | 47.46 | | Hellaswag (10-shot) | **78.55** | 76.11 | 76.13 | 76.91 | | Winogrande (5-shot) | 71.74 | **73.95** | 71.74 | 73.16 | | AGI Eval (0-shot) | 29.3 | **30.95** | 28.44 | 29.6 | | BigBench (0-shot) | 32.6 | 33.08 | 32.99 | **33.26** | | Average | 51.78 | **52.51** | 51.71 | 52.2 | ## Related Models | Model | Type | Data | Base Model | # Params | Download Links | |--------------------------|-----------------------------|-------------------|----------------------|------|------------------------------------------------------------------------| | Tamil LLaMA 7B v0.1 Base | Base model | 12GB | LLaMA 7B | 7B | [HF Hub](https://huggingface.co/abhinand/tamil-llama-7b-base-v0.1) | | Tamil LLaMA 13B v0.1 Base | Base model | 4GB | LLaMA 13B | 13B | [HF Hub](https://huggingface.co/abhinand/tamil-llama-13b-base-v0.1) | | Tamil LLaMA 7B v0.1 Instruct | Instruction following model | 145k instructions | Tamil LLaMA 7B Base | 7B | [HF Hub](https://huggingface.co/abhinand/tamil-llama-7b-instruct-v0.1) | | Tamil LLaMA 13B v0.1 Instruct | Instruction following model | 145k instructions | Tamil LLaMA 13B Base | 13B | [HF Hub](abhinand/tamil-llama-13b-instruct-v0.1) | | Tamil LLaMA 7B v0.2 Instruct | Instruction/Chat model | 500k instructions | Tamil LLaMA 7B Base v0.2 | 7B | [HF Hub](abhinand/tamil-llama-13b-instruct-v0.1) | | Malayalam LLaMA 7B v0.2 Instruct | Instruction/Chat model | 500k instructions | Malayalam LLaMA 7B Base v0.1 | 7B | [HF Hub](abhinand/tamil-llama-13b-instruct-v0.1) | ## Example Usage ```python from transformers import LlamaForCausalLM, AutoTokenizer, pipeline model = LlamaForCausalLM.from_pretrained( "abhinand/telugu-llama-instruct-v0.1", #load_in_8bit=True, # Set this depending on the GPU you have torch_dtype=torch.bfloat16, device_map={"": 0}, # Set this depending on the number of GPUs you have local_files_only=False # Optional ) model.eval() tokenizer = AutoTokenizer.from_pretrained("abhinand/telugu-llama-instruct-v0.1") inf_pipeline = pipeline("conversational", model=model, tokenizer=tokenizer) def format_instruction(system_prompt, question, return_dict=False): if system_prompt is None: messages = [ {'content': question, 'role': 'user'}, ] else: messages = [ {'content': system_prompt, 'role': 'system'}, {'content': question, 'role': 'user'}, ] if return_dict: return messages prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) return prompt # Set the generation configuration according to your needs temperature = 0.6 repetition_penalty = 1.1 max_new_tokens = 256 SYSTEM_PROMPT = "You are an AI assistant who follows instructions extremely well. Do your best your best to help." INPUT = "Who were the Nizams of Hyderabad?" instruction = format_instruction( system_prompt=SYSTEM_PROMPT, question=INPUT, return_dict=True, ) output = inf_pipeline( instruction, temperature=temperature, max_new_tokens=max_new_tokens, repetition_penalty=repetition_penalty ) print(output) ``` **Example Output:** ``` Conversation id: d57cdf33-01ff-4328-8efe-5c4fefdd6e77 system: You are an AI assistant who follows instructions extremely well. Do your best your best to help. user: Who were the Nizams of Hyderabad? assistant: The Nizams of Hyderabad were a dynasty that ruled the Deccan Plateau in southern India, including the city of Hyderabad. They were known for their wealth and patronage of art and culture. The last Nizam, Mir Osman Ali Khan, was one of the richest people in the world at the time of his death in 1967. ``` ## Usage Note It's important to note that the models have not undergone detoxification/censorship. Therefore, while they possess impressive linguistic capabilities, there is a possibility for them to generate content that could be deemed harmful or offensive. We urge users to exercise discretion and supervise the model's outputs closely, especially in public or sensitive applications. ## Meet the Developers Get to know the creators behind this innovative model and follow their contributions to the field: - [Abhinand Balachandran](https://www.linkedin.com/in/abhinand-05/) ## Citation If you use this model or any of the the Tamil-Llama related work in your research, please cite: ```bibtex @misc{balachandran2023tamilllama, title={Tamil-Llama: A New Tamil Language Model Based on Llama 2}, author={Abhinand Balachandran}, year={2023}, eprint={2311.05845}, archivePrefix={arXiv}, primaryClass={cs.CL} } ``` We hope this model serves as a valuable tool in your NLP toolkit and look forward to seeing the advancements it will enable in the understanding and generation of the Tamil language. # [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_abhinand__telugu-llama-7b-instruct-v0.1) | Metric |Value| |---------------------------------|----:| |Avg. |39.77| |AI2 Reasoning Challenge (25-Shot)|37.12| |HellaSwag (10-Shot) |67.92| |MMLU (5-Shot) |23.12| |TruthfulQA (0-shot) |49.05| |Winogrande (5-shot) |61.40| |GSM8k (5-shot) | 0.00|