abhinand's picture
Adding Evaluation Results (#2)
663842a verified
metadata
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 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. 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. Your support is invaluable and greatly appreciated.

"Buy Me A Coffee"

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 (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 on an RTX 3090 on runpod.

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 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
Tamil LLaMA 13B v0.1 Base Base model 4GB LLaMA 13B 13B HF Hub
Tamil LLaMA 7B v0.1 Instruct Instruction following model 145k instructions Tamil LLaMA 7B Base 7B HF Hub
Tamil LLaMA 13B v0.1 Instruct Instruction following model 145k instructions Tamil LLaMA 13B Base 13B HF Hub
Tamil LLaMA 7B v0.2 Instruct Instruction/Chat model 500k instructions Tamil LLaMA 7B Base v0.2 7B HF Hub
Malayalam LLaMA 7B v0.2 Instruct Instruction/Chat model 500k instructions Malayalam LLaMA 7B Base v0.1 7B HF Hub

Example Usage

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:

Citation

If you use this model or any of the the Tamil-Llama related work in your research, please cite:

@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

Detailed results can be found here

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