MicroLlama-300M

As an individual with limited access and compute, I have been wondering if I could build a decent large-language model for a while. As the big mega corporations are focused on getting bigger and bigger models, I am going small!

As a result, I set up the following goals to pretraining a 300M Llama model with the following restrictions:

1. My overall budget is $500.
2. Must pretrain an LLM from scratch with a fully open-source dataset and model.
3. Not allowed to finetune a model or use another LLM such as GPT-4 to generate any training data.

This project is heavily based on TinyLlama, which is an awesome open-source project aimed to pretraining a 1.1.1B Llama model on 1T tokens.

This project is work in progress. Currently, I have spent $280 on compute using 4 x Nvidia 4090 on Vast.ai and $3 on AWS S3 storage after 4 days of training of the 300M Llama model with 50B tokens.

I modified TinyLlama to support the following features (I will release my forked version of the source code after some clean up):

1. Pretrain a smaller size 300M model on Slimpajama
2. Removed Starcoderdata so that my model can focus on Slimpajama. This also means my model probably cannot do coding without fine-tuning.
3. Added the ability to process and tokenize Slimpajama while downloading the data. The original setup only works with pre-downloaded data. This turns out to be a good time-saver because downloading 800G+ of data on a non-commercial Internet is very slow, and processing all of Slimpajama data also takes time.
4. Various helper scripts and Python code such as python code for uploading the pretrained checkpoint to the huggingface hub.
5. Bug fixes.

Evaluation results

I performed the experiment using the standard lm-evaluation-harness setup. Following the same setup as TinyLlama, I used acc_norm for all datasets except for winogrande and boolq which used acc as the metrics.

1. keeeeenw/MicroLlama is the evaluation results for my 300M Llama model on 50B tokens.
2. google-best/bert-large-uncased is the baseline because it is one of the most popular small LLMs and it has a similar parameter count of 336M.
3. PY007/TinyLlama-1.1B-Chat-v0.1 as a sanity check I perform evaluation against one of the TinyLlama models to validate my setup for lm-evaluation-harness. These numbers are exactly the same as the ones reported by TinyLlama.
4. TinyLlama-1.1B-intermediate-step-1431k-3T is evaluation result for the best model created and reported by TinyLlama.

Model	Pretrain Tokens	HellaSwag	Obqa	WinoGrande	ARC_c	ARC_e	boolq	piqa	avg
keeeeenw/MicroLlama	50B	34.30	30.60	51.54	23.29	39.06	53.15	64.58	42.36
google-best/bert-large-uncased	N/A	24.53	26.20	49.80	25.68	25.08	40.86	47.66	34.26
PY007/TinyLlama-1.1B-Chat-v0.1	503B	53.81	32.20	55.01	28.67	49.62	58.04	69.64	49.57
TinyLlama-1.1B-intermediate-step-1431k-3T	3T	59.20	36.00	59.12	30.12	55.25	57.83	73.29	52.99

To reproduce my numbers, please install lm-evaluation-harness and run the following command:

lm_eval \
    --model hf \
    --model_args pretrained=keeeeenw/MicroLlama,dtype="float",tokenizer=TinyLlama/TinyLlama-1.1B-step-50K-105b \
    --tasks hellaswag,openbookqa,winogrande,arc_easy,arc_challenge,boolq,piqa \
    --device cuda:0 \
    --batch_size 64

Observations

1. Because keeeeenw/MicroLlama is much smaller than TinyLlama, our model does not achieve the same impressive results but the numbers are closer than I expected.
2. Our model outperforms google-best/bert-large-uncased which is actually slightly larger. The only dataset that google-best/bert-large-uncased outperformed our model is ARC_c (arc_challenge). I will provide more analysis as future study.

Based on the evaluation above, our model should be a good starting point for fine-tunning tasks that are typically performed using the BERT family of models. Some of tasks may include

1. (sentence transformer)[https://huggingface.co/sentence-transformers],
2. (bertscore)[https://huggingface.co/spaces/evaluate-metric/bertscore]
3. A light-weight chatbot after some finetuning.

Want to try it out?

1. Install dependencies

pip install transformers

2. Run code!

import torch
import transformers
from transformers import AutoTokenizer, LlamaForCausalLM

def generate_text(prompt, model, tokenizer):
    text_generator = transformers.pipeline(
        "text-generation",
        model=model,
        torch_dtype=torch.float16,
        device_map="auto",
        tokenizer=tokenizer
    )

    formatted_prompt = f"Question: {prompt} Answer:"

    sequences = text_generator(
        formatted_prompt,
        do_sample=True,
        top_k=5,
        top_p=0.9,
        num_return_sequences=1,
        repetition_penalty=1.5,
        max_new_tokens=128,
    )

    for seq in sequences:
        print(f"Result: {seq['generated_text']}")

# use the same tokenizer as TinyLlama
tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-step-50K-105b")

# load model from huggingface
# question from https://www.reddit.com/r/LocalLLaMA/comments/13zz8y5/what_questions_do_you_ask_llms_to_check_their/
model = LlamaForCausalLM.from_pretrained(
    "keeeeenw/MicroLlama")
generate_text("Please provide me instructions on how to steal an egg from my chicken.", model, tokenizer)

Acknowledgements

This repository is built upon TinyLlama which is based on lit-gpt and flash-attention.

@misc{zhang2024tinyllama,
      title={TinyLlama: An Open-Source Small Language Model}, 
      author={Peiyuan Zhang and Guangtao Zeng and Tianduo Wang and Wei Lu},
      year={2024},
      eprint={2401.02385},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
@online{lit-gpt,
  author    = {Lightning AI},
  title     = {Lit-GPT},
  url       = {https://github.com/Lightning-AI/lit-gpt},
  year      = {2023},
}
@article{dao2023flashattention2,
  title     ={Flash{A}ttention-2: Faster Attention with Better Parallelism and Work Partitioning},
  author    ={Dao, Tri},
  year      ={2023}
}